17 / app.py
razapro857's picture
Upload 7 files
d58eba4 verified
Raw
History Blame Contribute Delete
100 kB
import os, io, asyncio, tempfile, threading, re, subprocess, shutil, logging, secrets, sys, platform, hashlib, time, json
from contextlib import asynccontextmanager, nullcontext
try:
import firebase_admin
from firebase_admin import auth as fb_auth, firestore, credentials
HAS_FIREBASE = True
except Exception:
HAS_FIREBASE = False
firebase_admin = None
fb_auth = None
firestore = None
credentials = None
def _apply_space_config_secret():
"""Allow one protected HF secret to populate the usual env settings."""
raw = os.environ.get("VOICECRAFT_SPACE_CONFIG_JSON", "").strip()
if not raw:
return
try:
config = json.loads(raw)
if not isinstance(config, dict):
raise ValueError("VOICECRAFT_SPACE_CONFIG_JSON must be a JSON object")
except Exception as exc:
print("Space config JSON notice:", exc)
return
def set_default(env_name, *keys, transform=str):
if os.environ.get(env_name):
return
for key in keys:
if key in config and config[key] not in (None, ""):
os.environ[env_name] = transform(config[key])
return
firebase_json = (
config.get("firebase_service_account_json")
or config.get("firebase_admin_json")
or config.get("firebase_service_account")
or config.get("service_account")
)
if firebase_json and not os.environ.get("FIREBASE_SERVICE_ACCOUNT_JSON"):
os.environ["FIREBASE_SERVICE_ACCOUNT_JSON"] = (
firebase_json if isinstance(firebase_json, str) else json.dumps(firebase_json)
)
set_default("VOICECRAFT_SPACE_ID", "space_id", "voicecraft_space_id")
set_default("VOICECRAFT_AUTH_MODE", "auth_mode", "voicecraft_auth_mode")
set_default("VOICECRAFT_MAX_CONCURRENT_JOBS", "max_concurrent_jobs", "voicecraft_max_concurrent_jobs")
set_default("VOICECRAFT_SERVICE_VERSION", "service_version", "voicecraft_service_version")
set_default("MIN_CLIENT_VERSION", "min_client_version", "minimum_version")
set_default("ENABLE_CLONE_ENGINES", "enable_clone_engines", "clone_enabled", transform=lambda value: "true" if bool(value) else "false")
set_default("HF_TOKEN", "hf_token", "huggingface_token")
set_default("VOICECRAFT_API_SECRET", "api_secret", "voicecraft_api_secret")
_apply_space_config_secret()
_fb_app = None
_firestore_db = None
def _init_firebase():
global _fb_app, _firestore_db
if not HAS_FIREBASE:
return None
if _fb_app is None:
try:
credential_json = os.environ.get("FIREBASE_SERVICE_ACCOUNT_JSON", "").strip()
cred_path = os.environ.get('GOOGLE_APPLICATION_CREDENTIALS', '').strip()
if credential_json:
cred = credentials.Certificate(json.loads(credential_json))
_fb_app = firebase_admin.initialize_app(cred)
_firestore_db = firestore.client()
elif cred_path and os.path.exists(cred_path):
cred = credentials.Certificate(cred_path)
_fb_app = firebase_admin.initialize_app(cred)
_firestore_db = firestore.client()
except Exception as e:
print("Firebase init notice:", e)
return None
return _firestore_db
import requests
# Hide console window on Windows
CREATE_NO_WINDOW = 0x08000000 if sys.platform == "win32" else 0
from fastapi import FastAPI, Form, Request, HTTPException, UploadFile, File
from fastapi.responses import StreamingResponse, JSONResponse
from fastapi.middleware.gzip import GZipMiddleware
SERVICE_VERSION = os.environ.get("VOICECRAFT_SERVICE_VERSION", "3.1.2").strip()
_BACKGROUND_WARMUP_LOCK = threading.Lock()
_BACKGROUND_WARMUP_STARTED = False
@asynccontextmanager
async def _service_lifespan(_app):
"""Start model warm-up only when the ASGI service actually starts.
The old build launched download/model threads while merely importing this
module. That made tests, health tooling, and multi-worker process startup
unpredictable. A single guarded startup pass is enough for each process.
"""
global _BACKGROUND_WARMUP_STARTED
disabled = os.environ.get("VOICECRAFT_DISABLE_WARMUP", "").strip().lower() in {
"1", "true", "yes", "on"
}
if not disabled:
with _BACKGROUND_WARMUP_LOCK:
should_start = not _BACKGROUND_WARMUP_STARTED
_BACKGROUND_WARMUP_STARTED = True
if should_start:
threading.Thread(target=setup_piper, daemon=True, name="voicecraft-piper-warmup").start()
threading.Thread(target=setup_silero, daemon=True, name="voicecraft-silero-warmup").start()
if clone_engines_enabled():
threading.Thread(target=_warm_pocket_model, daemon=True, name="voicecraft-clone-warmup").start()
yield
app = FastAPI(
debug=False,
title="VoiceCraft TTS Server",
description="VoiceCraft desktop service.",
version=SERVICE_VERSION,
docs_url=None,
redoc_url=None,
openapi_url=None,
lifespan=_service_lifespan,
)
app.add_middleware(GZipMiddleware, minimum_size=1000)
# -----------------------------------------------------------------------------
# SECURITY — server-side Firebase authorization
# -----------------------------------------------------------------------------
_RAW_API_SECRET = os.environ.get("API_SECRET", "").strip()
if _RAW_API_SECRET.startswith("hf_") and not (
os.environ.get("HF_TOKEN")
or os.environ.get("HUGGINGFACE_HUB_TOKEN")
or os.environ.get("HUGGING_FACE_HUB_TOKEN")
):
os.environ["HF_TOKEN"] = _RAW_API_SECRET
API_SECRET = (
os.environ.get("VOICECRAFT_API_SECRET", "").strip()
or os.environ.get("APP_API_SECRET", "").strip()
or ("" if _RAW_API_SECRET.startswith("hf_") else _RAW_API_SECRET)
)
AUTH_MODE = os.environ.get("VOICECRAFT_AUTH_MODE", "license").strip().lower()
if AUTH_MODE not in {"license", "layered", "api_secret"}:
AUTH_MODE = "license"
MIN_CLIENT_VERSION = os.environ.get("MIN_CLIENT_VERSION", "3.0.0").strip()
MAX_CLONE_CHARACTERS = 60_000
CLONE_CHUNK_CHARACTERS = 1800
CLONE_RETRY_CHUNK_CHARACTERS = 900
CLONE_ENGINE_PREFIXES = ("f5tts:",)
CLONE_ENGINES = ["f5tts"]
CLONE_BACKEND_NAME = "pocket-tts-cpu"
BASE_ENGINES = ["edge", "piper", "silero"]
def _derive_space_id():
explicit = os.environ.get("VOICECRAFT_SPACE_ID", "").strip()
hf_space = os.environ.get("SPACE_ID", "").strip()
if explicit.lower() in {"auto", "hf", "huggingface"}:
explicit = ""
selected = explicit or hf_space or os.environ.get("SPACE_HOST", "").strip()
selected = selected.rsplit("/", 1)[-1].strip()
selected = re.sub(r"[^0-9A-Za-z_-]+", "-", selected).strip("-_")
return selected[:80] or "unregistered"
SPACE_ID = _derive_space_id()
def _derive_self_url():
"""Best-effort public URL of this Space, for access checks by URL."""
explicit = os.environ.get("VOICECRAFT_SPACE_URL", "").strip()
if explicit:
return explicit
host = os.environ.get("SPACE_HOST", "").strip()
if host:
return host if host.startswith("http") else f"https://{host}"
repo = os.environ.get("SPACE_ID", "").strip()
if "/" in repo:
owner, name = repo.split("/", 1)
slug = f"{owner}-{name}".replace("_", "-").replace(".", "-").lower()
return f"https://{slug}.hf.space"
return ""
SELF_SPACE_URL = _derive_self_url()
MAX_CONCURRENT_JOBS = max(1, min(int(os.environ.get("VOICECRAFT_MAX_CONCURRENT_JOBS", "1")), 20))
_SPACE_CAPACITY = threading.BoundedSemaphore(MAX_CONCURRENT_JOBS)
_SPACE_STATE_LOCK = threading.Lock()
_ACTIVE_JOBS = 0
_POLICY_LOCK = threading.Lock()
_POLICY_CACHE = {"loaded_at": 0.0, "maintenance_mode": False, "minimum_version": MIN_CLIENT_VERSION}
def _env_flag(name: str, default=None):
raw = os.environ.get(name)
if raw is None or raw == "":
return default
return raw.strip().lower() in {"1", "true", "yes", "on", "active", "enabled"}
def clone_engines_enabled() -> bool:
explicit = _env_flag("ENABLE_CLONE_ENGINES", None)
if explicit is not None:
return explicit
# CPU clone is now the default backend. Set ENABLE_CLONE_ENGINES=0 for TTS-only Spaces.
return True
def clone_backend_available() -> bool:
if not clone_engines_enabled():
return False
if _POCKET_MODEL_ERROR or _POCKET_GENERATION_ERROR:
return False
if _POCKET_MODEL is not None:
return bool(getattr(_POCKET_MODEL, "has_voice_cloning", False))
return _hf_token_configured()
def clone_backend_status() -> dict:
if not clone_engines_enabled():
return {"enabled": False, "ready": False, "message": "disabled"}
if _POCKET_GENERATION_ERROR:
return {"enabled": False, "ready": False, "message": "generation_failed"}
if _POCKET_MODEL is not None:
ready = bool(getattr(_POCKET_MODEL, "has_voice_cloning", False))
return {
"enabled": ready,
"ready": ready,
"message": "ready" if ready else "model_loaded_without_clone_weights",
}
if _POCKET_MODEL_ERROR:
return {"enabled": False, "ready": False, "message": "model_startup_failed"}
if not _hf_token_configured():
return {"enabled": False, "ready": False, "message": "hf_token_required"}
return {"enabled": True, "ready": False, "message": "loading"}
def available_engines():
return BASE_ENGINES + (CLONE_ENGINES if clone_backend_available() else [])
def get_runtime_policy():
now = time.monotonic()
with _POLICY_LOCK:
if now - float(_POLICY_CACHE.get("loaded_at", 0.0)) < 30:
return dict(_POLICY_CACHE)
policy = {
"loaded_at": now,
"maintenance_mode": False,
"minimum_version": MIN_CLIENT_VERSION,
"space_enabled": True,
"space_clone_enabled": clone_backend_available(),
}
db = _init_firebase()
if db is not None:
try:
runtime_doc = db.collection("public_config").document("runtime").get()
if runtime_doc.exists:
runtime = runtime_doc.to_dict() or {}
policy["maintenance_mode"] = bool(runtime.get("maintenance_mode", False))
policy["minimum_version"] = str(runtime.get("minimum_version") or MIN_CLIENT_VERSION)
if SPACE_ID != "unregistered":
space_doc = db.collection("spaces").document(SPACE_ID).get()
if space_doc.exists:
space_data = space_doc.to_dict() or {}
policy["space_enabled"] = bool(space_data.get("enabled", False))
policy["space_clone_enabled"] = bool(space_data.get("clone_enabled", False)) and clone_backend_available()
except Exception:
pass
with _POLICY_LOCK:
_POLICY_CACHE.clear()
_POLICY_CACHE.update(policy)
return dict(policy)
def _clean_hf_space_url(value):
raw = str(value or "").strip().rstrip("/")
if not raw.startswith("https://"):
return ""
host = raw.split("/", 3)[2].lower()
if host == "hf.space" or host.endswith(".hf.space") or host == "huggingface.co" or host.endswith(".huggingface.co"):
return raw
return ""
def _space_public_payload(snapshot_id, item):
url = _clean_hf_space_url((item or {}).get("url"))
if not url:
return None
return {
"id": str(snapshot_id or "")[:80],
"name": str((item or {}).get("name") or snapshot_id or "")[:80],
"url": url,
# Default False, matching get_runtime_policy() and the admin panel's
# space_row(). It used to default True here, so a Space document with
# no clone_enabled field advertised cloning to the desktop and then
# refused every clone job with a 403 from the policy check.
"clone_enabled": bool((item or {}).get("clone_enabled", (item or {}).get("clone", False))),
"priority": int((item or {}).get("priority", 100) or 100),
"max_concurrent_jobs": max(1, int((item or {}).get("max_concurrent_jobs", 1) or 1)),
}
def _license_space_pool(db, uid, user_data, license_key, lic):
"""Return only the Spaces this signed-in user may see/use.
Access is decided by ``space_access_mode`` on the user document:
pool assigned Spaces first, then the shared public pool
assigned_only ONLY the Spaces the admin ticked for this user
blocked nothing at all
Private Spaces (``visibility == "private"``) never enter the shared pool;
they are handed out exclusively to the users they were assigned to.
"""
if db is None:
return []
assigned_ids = set()
assigned_urls = set()
assigned_tts_ids = set()
assigned_tts_urls = set()
assigned_clone_ids = set()
assigned_clone_urls = set()
mode = ""
for source in (user_data or {}, lic or {}):
if not mode:
mode = str(source.get("space_access_mode") or "").strip().lower()
for key in ("space_ids", "clone_space_ids", "dedicated_space_ids", "assigned_space_ids"):
for value in source.get(key) or []:
if value:
assigned_ids.add(str(value).strip())
for key in ("spaces", "clone_spaces", "dedicated_spaces", "assigned_spaces"):
for value in source.get(key) or []:
if isinstance(value, dict):
payload = _space_public_payload(value.get("id") or value.get("name"), value)
if payload:
assigned_urls.add(payload["url"])
else:
url = _clean_hf_space_url(value)
if url:
assigned_urls.add(url)
else:
assigned_ids.add(str(value).strip())
for key in ("assigned_tts_space_ids",):
for value in source.get(key) or []:
if value:
assigned_tts_ids.add(str(value).strip())
for key in ("assigned_clone_space_ids",):
for value in source.get(key) or []:
if value:
assigned_clone_ids.add(str(value).strip())
for value in source.get("assigned_tts_spaces") or []:
url = _clean_hf_space_url(value)
if url:
assigned_tts_urls.add(url)
elif value:
assigned_tts_ids.add(str(value).strip())
for value in source.get("assigned_clone_spaces") or []:
url = _clean_hf_space_url(value)
if url:
assigned_clone_urls.add(url)
elif value:
assigned_clone_ids.add(str(value).strip())
if mode not in {"pool", "assigned_only", "blocked"}:
mode = "pool"
if mode == "blocked":
return []
assigned_ids |= assigned_tts_ids | assigned_clone_ids
assigned_urls |= assigned_tts_urls | assigned_clone_urls
dedicated = []
dedicated_unhealthy = []
auto_pool = []
try:
for snapshot in db.collection("spaces").where("enabled", "==", True).stream():
item = snapshot.to_dict() or {}
# Health synchronization marks paused/crashed workers false. Those
# endpoints are kept out of the shared pool so the desktop doesn't
# waste time failing over from known-dead Spaces. A dedicated
# Space is different: for an assigned_only customer it may be the
# only route they have, and a stale flag from one failed probe
# would otherwise take them fully offline. Those are held back as
# a last resort instead of being dropped.
space_unhealthy = item.get("last_health_ok") is False
payload = _space_public_payload(snapshot.id, item)
if not payload:
continue
visibility = str(item.get("visibility") or "").strip().lower()
if visibility not in {"public", "private"}:
visibility = "private" if (
item.get("assigned_uid") or item.get("assigned_license_key")
) else "public"
has_scoped_assignments = bool(
assigned_tts_ids or assigned_tts_urls
or assigned_clone_ids or assigned_clone_urls
)
legacy_direct_match = (
str(item.get("assigned_uid") or "").strip() == str(uid or "").strip()
or str(
item.get("assigned_license_key") or item.get("assigned_license") or ""
).strip() == str(license_key or "").strip()
)
is_dedicated = (
snapshot.id in assigned_ids
or payload["url"] in assigned_urls
or (legacy_direct_match and not has_scoped_assignments)
)
if is_dedicated:
# Per-Space entitlements: a worker assigned for TTS only must
# not accept clone jobs from this user, and vice versa.
tts_ok = True
clone_ok = bool(payload["clone_enabled"])
if assigned_tts_ids or assigned_tts_urls or assigned_clone_ids or assigned_clone_urls:
tts_ok = (
snapshot.id in assigned_tts_ids
or payload["url"] in assigned_tts_urls
)
clone_ok = clone_ok and (
snapshot.id in assigned_clone_ids
or payload["url"] in assigned_clone_urls
)
if not tts_ok and not clone_ok and not has_scoped_assignments:
# Only present through a legacy combined list.
tts_ok = True
payload = dict(payload, tts_enabled=tts_ok, clone_enabled=clone_ok, dedicated=True)
(dedicated_unhealthy if space_unhealthy else dedicated).append(payload)
elif (
not space_unhealthy
and visibility == "public"
and not item.get("assigned_uid")
and not item.get("assigned_license_key")
and not item.get("assigned_license")
):
auto_pool.append(dict(payload, tts_enabled=True, dedicated=False))
except Exception:
return []
if mode == "assigned_only":
auto_pool = []
if not dedicated and not auto_pool and dedicated_unhealthy:
# Better to hand back a Space that failed its last probe than to hand
# back nothing: the desktop re-checks /health itself before using one.
dedicated = dedicated_unhealthy
dedicated.sort(key=lambda item: (item["priority"], item["name"].lower()))
auto_pool.sort(key=lambda item: (item["priority"], item["name"].lower()))
seen_urls = set()
rows = []
for item in dedicated + auto_pool:
if item["url"] not in seen_urls:
rows.append(item)
seen_urls.add(item["url"])
return rows
def _space_pool_is_exclusive(user_data, lic):
"""True when the desktop must replace its cached pool, not merge into it."""
for source in (user_data or {}, lic or {}):
mode = str(source.get("space_access_mode") or "").strip().lower()
if mode in {"assigned_only", "blocked"}:
return True
if mode == "pool":
return False
return False
def verify_token(request: Request):
if not API_SECRET:
logging.warning("API_SECRET missing in Hugging Face Space secrets")
raise HTTPException(status_code=503, detail="Service unavailable")
token = request.headers.get("X-API-Token", "").strip()
if not secrets.compare_digest(token, API_SECRET):
raise HTTPException(status_code=403, detail="Unauthorized")
_LICENSE_RATE_STATE = {}
_LICENSE_RATE_LOCK = threading.Lock()
def _bounded_env_int(name: str, default: int, minimum: int, maximum: int) -> int:
try:
return max(minimum, min(maximum, int(os.environ.get(name, default))))
except (TypeError, ValueError):
return default
def _version_tuple(value: str) -> tuple[int, ...]:
parts = re.findall(r"\d+", str(value or ""))
numbers = [int(part) for part in parts[:4]]
return tuple((numbers + [0, 0, 0, 0])[:4])
def _coerce_utc_datetime(value):
"""Parse Firestore timestamps and ISO strings into aware UTC datetimes."""
from datetime import date, datetime, timezone
if value in (None, ""):
return None
if isinstance(value, datetime):
parsed = value
elif isinstance(value, date):
parsed = datetime(value.year, value.month, value.day)
elif isinstance(value, str):
raw = value.strip()
if not raw:
return None
try:
parsed = datetime.fromisoformat(raw.replace("Z", "+00:00"))
except ValueError:
return None
else:
return None
if parsed.tzinfo is None:
parsed = parsed.replace(tzinfo=timezone.utc)
return parsed.astimezone(timezone.utc)
def _require_firebase_db():
"""Return a usable Firebase client or a truthful 503 response."""
if not HAS_FIREBASE or fb_auth is None or firestore is None:
raise HTTPException(503, "Firebase authentication is not configured")
db = _init_firebase()
if db is None:
raise HTTPException(503, "Firebase database is unavailable")
return db
def _reserve_monthly_usage(db, uid, license_key, lic, is_clone, requested_characters):
from datetime import datetime, timezone
requested_characters = max(0, int(requested_characters or 0))
if not requested_characters:
return
month_id = datetime.now(timezone.utc).strftime("%Y%m")
usage_ref = db.collection("usage_monthly").document(f"{uid}_{month_id}")
used_field = "clone_characters" if is_clone else "tts_characters"
reserved_field = "clone_reserved" if is_clone else "tts_reserved"
limit = int(
lic.get("monthly_clone_characters" if is_clone else "monthly_characters")
or 0
)
transaction = db.transaction()
@firestore.transactional
def reserve(transaction):
snapshot = usage_ref.get(transaction=transaction)
usage = snapshot.to_dict() if snapshot.exists else {}
used = max(0, int(usage.get(used_field) or 0))
reserved = max(0, int(usage.get(reserved_field) or 0))
if limit and used + reserved + requested_characters > limit:
label = "voice-clone" if is_clone else "TTS character"
raise HTTPException(
status_code=403,
detail=f"Monthly {label} allowance has been reached",
)
transaction.set(
usage_ref,
{
"uid": uid,
"license_key": license_key,
"period": month_id,
reserved_field: reserved + requested_characters,
"updated_at": firestore.SERVER_TIMESTAMP,
},
merge=True,
)
try:
reserve(transaction)
except HTTPException:
raise
except Exception as exc:
logging.warning("Usage reservation failed: %s", exc.__class__.__name__)
raise HTTPException(503, "Usage allowance could not be verified")
def _release_monthly_reservation(uid, is_clone, characters):
"""Give a reservation back when the request never reaches record_usage.
``_reserve_monthly_usage`` books characters against the customer's monthly
allowance before generation starts, and ``record_usage`` returns them in a
``finally`` block. Anything that rejects the request *between* those two
points (rate limiting, a per-request size check) would otherwise leak the
reservation permanently, slowly eating the customer's paid allowance until
an admin resets it by hand.
"""
characters = max(0, int(characters or 0))
if not characters or not uid:
return
from datetime import datetime, timezone
db = _init_firebase()
if db is None:
return
reserved_field = "clone_reserved" if is_clone else "tts_reserved"
month_id = datetime.now(timezone.utc).strftime("%Y%m")
usage_ref = db.collection("usage_monthly").document(f"{uid}_{month_id}")
transaction = db.transaction()
@firestore.transactional
def release(transaction):
snapshot = usage_ref.get(transaction=transaction)
usage = snapshot.to_dict() if snapshot.exists else {}
current = max(0, int(usage.get(reserved_field) or 0))
if current <= 0:
return
transaction.set(
usage_ref,
{
reserved_field: max(0, current - characters),
"updated_at": firestore.SERVER_TIMESTAMP,
},
merge=True,
)
try:
release(transaction)
except Exception as exc:
logging.warning("Reservation release failed: %s", exc.__class__.__name__)
def _monthly_credit_summary(db, uid, lic):
"""Return the current UTC month's enforced credit balance for the client."""
from datetime import datetime, timezone
month_id = datetime.now(timezone.utc).strftime("%Y%m")
snapshot = db.collection("usage_monthly").document(f"{uid}_{month_id}").get()
usage = snapshot.to_dict() if snapshot.exists else {}
def balance(limit_field, used_field, reserved_field):
limit = max(0, int(lic.get(limit_field) or 0))
used = max(0, int(usage.get(used_field) or 0))
reserved = max(0, int(usage.get(reserved_field) or 0))
unlimited = limit == 0
remaining = None if unlimited else max(0, limit - used - reserved)
return limit, used, reserved, remaining, unlimited
tts = balance("monthly_characters", "tts_characters", "tts_reserved")
clone = balance(
"monthly_clone_characters", "clone_characters", "clone_reserved"
)
return {
"usage_period": month_id,
"monthly_characters": tts[0],
"tts_characters_used": tts[1],
"tts_characters_reserved": tts[2],
"tts_characters_remaining": tts[3],
"tts_unlimited": tts[4],
"monthly_clone_characters": clone[0],
"clone_characters_used": clone[1],
"clone_characters_reserved": clone[2],
"clone_characters_remaining": clone[3],
"clone_unlimited": clone[4],
}
def _register_device_tx(db, user_ref, device_id, max_devices):
"""Transactionally register a device against the user's device list."""
max_devices = max(1, int(max_devices or 1))
@firestore.transactional
def register(transaction):
snapshot = user_ref.get(transaction=transaction)
current = snapshot.to_dict() if snapshot.exists else {}
ids = list(current.get("device_ids") or [])
if device_id not in ids:
if len(ids) >= max_devices:
raise HTTPException(403, "This device is not registered for the license")
ids.append(device_id)
transaction.update(user_ref, {"device_ids": ids})
return ids
try:
return register(db.transaction())
except HTTPException:
raise
except Exception as exc:
logging.warning("Device registration failed: %s", exc.__class__.__name__)
raise HTTPException(503, "Device registration could not be verified")
def _collect_space_assignments(user_data, lic, is_clone: bool):
"""Return request-scoped and all assignment references for one account."""
sources = (user_data or {}, lic or {})
scoped_id_key = "assigned_clone_space_ids" if is_clone else "assigned_tts_space_ids"
scoped_url_key = "assigned_clone_spaces" if is_clone else "assigned_tts_spaces"
scoped_keys = (
"assigned_tts_space_ids", "assigned_clone_space_ids",
"assigned_tts_spaces", "assigned_clone_spaces",
)
has_scoped_lists = any(key in source for source in sources for key in scoped_keys)
request_ids, request_urls = set(), set()
all_ids, all_urls = set(), set()
for source in sources:
for key in ("assigned_tts_space_ids", "assigned_clone_space_ids", "assigned_space_ids"):
for value in source.get(key) or []:
if value:
all_ids.add(str(value).strip())
for key in ("assigned_tts_spaces", "assigned_clone_spaces", "assigned_spaces"):
for value in source.get(key) or []:
cleaned = _clean_hf_space_url(value)
if cleaned:
all_urls.add(cleaned)
id_keys = (scoped_id_key,) if has_scoped_lists else (scoped_id_key, "assigned_space_ids")
url_keys = (scoped_url_key,) if has_scoped_lists else (scoped_url_key, "assigned_spaces")
for key in id_keys:
for value in source.get(key) or []:
if value:
request_ids.add(str(value).strip())
for key in url_keys:
for value in source.get(key) or []:
cleaned = _clean_hf_space_url(value)
if cleaned:
request_urls.add(cleaned)
return request_ids, request_urls, all_ids, all_urls, has_scoped_lists
def _enforce_space_access(uid, user_data, lic, is_clone: bool, db=None, license_key=""):
"""Enforce blocked, scoped, assigned-only, and private-worker access.
The old implementation only checked assignments when a user was in
``assigned_only`` mode. A pool-mode customer who learned a private worker
URL could therefore call that worker directly, and a TTS-only assignment
could be used for cloning. Public pool workers remain available normally;
private/dedicated workers and scoped grants are now authoritative on every
request.
"""
mode = ""
for source in (user_data or {}, lic or {}):
candidate = str(source.get("space_access_mode") or "").strip().lower()
if candidate:
mode = candidate
break
if mode not in {"pool", "assigned_only", "blocked"}:
mode = "pool"
if mode == "blocked":
raise HTTPException(
403,
"Server access for this account has been disabled. Please contact support.",
)
request_ids, request_urls, all_ids, all_urls, has_scoped_lists = _collect_space_assignments(
user_data, lic, is_clone
)
own_url = _clean_hf_space_url(SELF_SPACE_URL)
assigned_for_request = SPACE_ID in request_ids or bool(own_url and own_url in request_urls)
listed_in_any_scope = SPACE_ID in all_ids or bool(own_url and own_url in all_urls)
# If this account explicitly lists the worker for one service only, the
# other service must not be able to use the same URL even in pool mode.
if has_scoped_lists and listed_in_any_scope and not assigned_for_request:
service = "voice cloning" if is_clone else "text-to-speech"
raise HTTPException(403, f"This server is not assigned for {service} on your account")
space_data = {}
if SPACE_ID != "unregistered":
db = db or _init_firebase()
if db is None:
raise HTTPException(503, "Server access policy is unavailable")
try:
snapshot = db.collection("spaces").document(SPACE_ID).get()
if snapshot.exists:
space_data = snapshot.to_dict() or {}
except HTTPException:
raise
except Exception as exc:
logging.warning("Space access lookup failed: %s", exc.__class__.__name__)
raise HTTPException(503, "Server access policy could not be verified")
visibility = str(space_data.get("visibility") or "").strip().lower()
directly_assigned_uid = str(space_data.get("assigned_uid") or "").strip()
directly_assigned_license = str(
space_data.get("assigned_license_key") or space_data.get("assigned_license") or ""
).strip()
legacy_direct_match = bool(
(directly_assigned_uid and directly_assigned_uid == str(uid or "").strip())
or (
directly_assigned_license
and directly_assigned_license == str(license_key or user_data.get("license_key") or "").strip()
)
)
is_reserved_worker = visibility == "private" or bool(
directly_assigned_uid or directly_assigned_license
)
if is_reserved_worker:
allowed = assigned_for_request or (legacy_direct_match and not has_scoped_lists)
if not allowed:
raise HTTPException(403, "This private server is not assigned to your account")
if mode == "assigned_only" and not (
assigned_for_request or (legacy_direct_match and not has_scoped_lists)
):
raise HTTPException(
403,
"This server is not assigned to your account. Restart VoiceCraft to "
"refresh your assigned servers, or contact support.",
)
def verify_firebase_auth(request, is_clone: bool = False, requested_characters: int = 0):
_require_firebase_db()
auth_header = request.headers.get("Authorization", "")
if not auth_header.startswith("Bearer "):
raise HTTPException(401, "Missing authentication token")
token = auth_header.split("Bearer ", 1)[1]
try:
decoded = fb_auth.verify_id_token(token)
except Exception:
raise HTTPException(401, "Invalid or expired authentication token")
uid = decoded["uid"]
db = _require_firebase_db()
user_doc = db.collection("users").document(uid).get()
if not user_doc.exists:
raise HTTPException(403, "User account not found")
user_data = user_doc.to_dict() or {}
if user_data.get("is_blocked"):
raise HTTPException(403, "Account has been blocked")
device_id = str(request.headers.get("X-Device-ID", "")).strip()[:128]
if not device_id:
raise HTTPException(403, "Device identification is required")
license_key = user_data.get("license_key")
if not license_key:
raise HTTPException(403, "No active license. Activate a license key from Account settings")
license_doc = db.collection("licenses").document(license_key).get()
if not license_doc.exists:
raise HTTPException(403, "License key not found")
lic = license_doc.to_dict()
if lic.get("status") not in ("active",):
raise HTTPException(403, f"License is {lic.get('status', 'invalid')}")
max_devs = max(1, int(lic.get("max_devices") or 1))
device_ids = _register_device_tx(db, user_doc.reference, device_id, max_devs)
if len(device_ids) > max_devs:
raise HTTPException(403, "Maximum licensed devices exceeded")
from datetime import datetime, timezone
raw_expiry = lic.get("expiry_date")
expiry = _coerce_utc_datetime(raw_expiry)
if raw_expiry not in (None, "") and expiry is None:
raise HTTPException(403, "License expiry is invalid. Please contact support")
if expiry and expiry < datetime.now(timezone.utc):
raise HTTPException(403, "License has expired")
if is_clone and (
not lic.get("voice_clone")
or not user_data.get("voice_clone_enabled", lic.get("voice_clone", False))
):
raise HTTPException(403, "Voice cloning not included in your plan")
# Server-side access enforcement. The desktop is told which Spaces it may
# use, but a customer who saved an old URL could still call this worker
# directly, so the worker checks for itself on every request.
_enforce_space_access(uid, user_data, lic, is_clone, db=db, license_key=license_key)
_reserve_monthly_usage(
db,
uid,
license_key,
lic,
is_clone,
requested_characters,
)
return uid, license_key, lic, user_data
def _enforce_rate_limit(uid: str, is_clone: bool):
now = time.monotonic()
window = 60.0
limit = _bounded_env_int(
"CLONE_REQUESTS_PER_MINUTE" if is_clone else "TTS_REQUESTS_PER_MINUTE",
8 if is_clone else 60,
1,
600,
)
state_key = f"{uid}:{'clone' if is_clone else 'tts'}"
with _LICENSE_RATE_LOCK:
requests_in_window = [
timestamp
for timestamp in _LICENSE_RATE_STATE.get(state_key, [])
if now - timestamp < window
]
if len(requests_in_window) >= limit:
raise HTTPException(status_code=429, detail="Request limit reached; try again shortly")
requests_in_window.append(now)
_LICENSE_RATE_STATE[state_key] = requests_in_window
async def authorize_request(request: Request, engine: str, requested_characters: int = 0) -> dict:
policy = await asyncio.to_thread(get_runtime_policy)
if policy.get("maintenance_mode"):
raise HTTPException(status_code=503, detail="VoiceCraft is temporarily under maintenance")
if not policy.get("space_enabled", True):
raise HTTPException(status_code=503, detail="This worker has been disabled")
if engine in CLONE_ENGINES and not policy.get("space_clone_enabled", False):
raise HTTPException(status_code=403, detail="Voice cloning is not available on this server")
client_version = request.headers.get("X-Client-Version", "").strip()
minimum_version = str(policy.get("minimum_version") or MIN_CLIENT_VERSION)
if client_version and _version_tuple(client_version) < _version_tuple(minimum_version):
raise HTTPException(status_code=426, detail="VoiceCraft update required")
if AUTH_MODE in {"api_secret", "layered"}:
verify_token(request)
if AUTH_MODE == "api_secret":
return {
"uid": "api_secret_system",
"license_key": "api_secret_key",
"license": {},
"usage_reserved": False,
}
if not client_version:
raise HTTPException(status_code=426, detail="VoiceCraft client version is required")
is_clone = engine in CLONE_ENGINES
uid, license_key, lic, _user = await asyncio.to_thread(
verify_firebase_auth,
request,
is_clone,
requested_characters,
)
try:
_enforce_rate_limit(uid, is_clone)
except Exception:
# The allowance was already booked inside verify_firebase_auth; a
# throttled request must not cost the customer any characters.
await asyncio.to_thread(
_release_monthly_reservation, uid, is_clone, requested_characters
)
raise
return {
"uid": uid,
"license_key": license_key,
"license": lic,
"usage_reserved": bool(requested_characters),
}
def acquire_worker_slot():
global _ACTIVE_JOBS
if not _SPACE_CAPACITY.acquire(blocking=False):
raise HTTPException(status_code=503, detail="Worker is busy; try another Space")
with _SPACE_STATE_LOCK:
_ACTIVE_JOBS += 1
active = _ACTIVE_JOBS
db = _init_firebase()
if db is not None and SPACE_ID != "unregistered":
try:
_policy = get_runtime_policy()
db.collection("spaces").document(SPACE_ID).set({
"active_jobs": active,
"last_heartbeat": firestore.SERVER_TIMESTAMP,
"enabled": bool(_policy.get("space_enabled", True)),
"clone_enabled": bool(_policy.get("space_clone_enabled", False)),
}, merge=True)
except Exception:
pass
def release_worker_slot():
global _ACTIVE_JOBS
with _SPACE_STATE_LOCK:
_ACTIVE_JOBS = max(0, _ACTIVE_JOBS - 1)
active = _ACTIVE_JOBS
_SPACE_CAPACITY.release()
db = _init_firebase()
if db is not None and SPACE_ID != "unregistered":
try:
_policy = get_runtime_policy()
db.collection("spaces").document(SPACE_ID).set({
"active_jobs": active,
"last_heartbeat": firestore.SERVER_TIMESTAMP,
"enabled": bool(_policy.get("space_enabled", True)),
"clone_enabled": bool(_policy.get("space_clone_enabled", False)),
}, merge=True)
except Exception:
pass
def record_usage(auth_context, engine, characters, status, duration_ms):
uid = str((auth_context or {}).get("uid") or "")
if not uid:
return
db = _init_firebase()
if db is None:
return
is_clone = engine in CLONE_ENGINES
event = {
"uid": uid,
"license_key": str(auth_context.get("license_key") or ""),
"space_id": SPACE_ID,
"engine": engine,
"characters": int(characters),
"is_clone": is_clone,
"status": status,
"duration_ms": int(duration_ms),
"created_at": firestore.SERVER_TIMESTAMP,
}
try:
batch = db.batch()
event_ref = db.collection("usage_events").document()
user_ref = db.collection("users").document(uid)
space_ref = db.collection("spaces").document(SPACE_ID)
from datetime import datetime, timezone
month_id = datetime.now(timezone.utc).strftime("%Y%m")
monthly_ref = db.collection("usage_monthly").document(f"{uid}_{month_id}")
batch.set(event_ref, event)
billed_characters = int(characters) if status == "success" else 0
reserved_field = "clone_reserved" if is_clone else "tts_reserved"
user_updates = {
"generation_count": firestore.Increment(1),
"last_generation_at": firestore.SERVER_TIMESTAMP,
"tts_characters": firestore.Increment(0 if is_clone else billed_characters),
"clone_characters": firestore.Increment(billed_characters if is_clone else 0),
}
batch.set(user_ref, user_updates, merge=True)
monthly_updates = {
"uid": uid,
"period": month_id,
"tts_characters": firestore.Increment(0 if is_clone else billed_characters),
"clone_characters": firestore.Increment(billed_characters if is_clone else 0),
"generation_count": firestore.Increment(1),
"updated_at": firestore.SERVER_TIMESTAMP,
}
if auth_context.get("usage_reserved"):
monthly_updates[reserved_field] = firestore.Increment(-int(characters))
batch.set(monthly_ref, monthly_updates, merge=True)
if SPACE_ID != "unregistered":
batch.set(space_ref, {
"total_requests": firestore.Increment(1),
"total_characters": firestore.Increment(billed_characters),
"last_heartbeat": firestore.SERVER_TIMESTAMP,
}, merge=True)
batch.commit()
except Exception as exc:
logging.warning("Usage logging failed: %s", exc.__class__.__name__)
# Firestore batches are atomic. If the batch failed, release the
# pre-authorized allowance so a transient logging fault cannot consume
# customer credits permanently.
if auth_context.get("usage_reserved"):
_release_monthly_reservation(uid, is_clone, characters)
# -----------------------------------------------------------------------------
# PIPER TTS SETUP - Auto download on first run
# -----------------------------------------------------------------------------
PIPER_DIR = "/tmp/piper"
PIPER_BIN = os.path.join(PIPER_DIR, "piper")
PIPER_MODELS_DIR = "/tmp/piper_models"
PIPER_READY = False
PIPER_VOICES = {
# English
"piper:en_US-amy-medium": ("en_US-amy-medium.onnx", "en_US-amy-medium.onnx.json"),
"piper:en_US-joe-medium": ("en_US-joe-medium.onnx", "en_US-joe-medium.onnx.json"),
"piper:en_US-lessac-medium": ("en_US-lessac-medium.onnx", "en_US-lessac-medium.onnx.json"),
"piper:en_US-ryan-high": ("en_US-ryan-high.onnx", "en_US-ryan-high.onnx.json"),
"piper:en_GB-alan-medium": ("en_GB-alan-medium.onnx", "en_GB-alan-medium.onnx.json"),
"piper:en_GB-alba-medium": ("en_GB-alba-medium.onnx", "en_GB-alba-medium.onnx.json"),
# Urdu / Hindi / Arabic
"piper:ur_PK-fasih-medium": ("ur_PK-fasih-medium.onnx", "ur_PK-fasih-medium.onnx.json"),
"piper:hi_IN-pratham-medium": ("hi_IN-pratham-medium.onnx", "hi_IN-pratham-medium.onnx.json"),
"piper:ar_JO-kareem-medium": ("ar_JO-kareem-medium.onnx", "ar_JO-kareem-medium.onnx.json"),
# Other languages
"piper:de_DE-thorsten-medium": ("de_DE-thorsten-medium.onnx", "de_DE-thorsten-medium.onnx.json"),
"piper:fr_FR-upmc-medium": ("fr_FR-upmc-medium.onnx", "fr_FR-upmc-medium.onnx.json"),
"piper:ru_RU-irina-medium": ("ru_RU-irina-medium.onnx", "ru_RU-irina-medium.onnx.json"),
"piper:tr_TR-dfki-medium": ("tr_TR-dfki-medium.onnx", "tr_TR-dfki-medium.onnx.json"),
"piper:pt_BR-faber-medium": ("pt_BR-faber-medium.onnx", "pt_BR-faber-medium.onnx.json"),
"piper:nl_NL-mls-medium": ("nl_NL-mls-medium.onnx", "nl_NL-mls-medium.onnx.json"),
}
PIPER_BASE_URL = "https://huggingface.co/rhasspy/piper-voices/resolve/main"
def setup_piper():
global PIPER_READY
try:
import platform
os.makedirs(PIPER_DIR, exist_ok=True)
os.makedirs(PIPER_MODELS_DIR, exist_ok=True)
system = platform.system().lower()
arch = platform.machine().lower()
if system == "linux" and "x86" in arch:
piper_url = "https://github.com/rhasspy/piper/releases/download/2023.11.14-2/piper_linux_x86_64.tar.gz"
elif system == "linux" and "aarch" in arch:
piper_url = "https://github.com/rhasspy/piper/releases/download/2023.11.14-2/piper_linux_aarch64.tar.gz"
else:
print(f"[Piper] unsupported platform {system}/{arch}, Piper disabled")
return
if not os.path.exists(PIPER_BIN):
print("[Piper] downloading binary...")
import urllib.request
tar_path = "/tmp/piper.tar.gz"
urllib.request.urlretrieve(piper_url, tar_path)
import tarfile
with tarfile.open(tar_path, "r:gz") as tf:
tf.extractall("/tmp/piper_extract")
extracted = "/tmp/piper_extract/piper"
if os.path.isdir(extracted):
for item in os.listdir(extracted):
shutil.move(os.path.join(extracted, item), os.path.join(PIPER_DIR, item))
else:
shutil.move(extracted, PIPER_BIN)
os.chmod(PIPER_BIN, 0o755)
print("[OK] Piper binary ready")
PIPER_READY = True
print("[OK] Piper TTS ready")
except Exception as e:
print(f"[WARN] Piper setup failed (non-critical): {e}")
PIPER_READY = False
def download_piper_model(voice_code: str) -> tuple:
"""Model yoksa indir, path tuple dondur (onnx, json)"""
if voice_code not in PIPER_VOICES:
raise ValueError(f"Unknown Piper voice: {voice_code}")
onnx_file, json_file = PIPER_VOICES[voice_code]
onnx_path = os.path.join(PIPER_MODELS_DIR, onnx_file)
json_path = os.path.join(PIPER_MODELS_DIR, json_file)
import urllib.request
# Build correct HF path: en/en_US/amy/medium/en_US-amy-medium.onnx
parts = onnx_file.rsplit("-", 2)
lang_code = parts[0] # en_US
voice = parts[1] # amy
quality = parts[2].replace(".onnx", "") # medium
lang_short = lang_code.split("_")[0] # en
hf_dir = f"{lang_short}/{lang_code}/{voice}/{quality}"
for fname, fpath in [(onnx_file, onnx_path), (json_file, json_path)]:
if not os.path.exists(fpath):
url = f"{PIPER_BASE_URL}/{hf_dir}/{fname}"
print(f"[DOWNLOAD] Downloading Piper model: {fname}")
try:
urllib.request.urlretrieve(url, fpath)
except Exception:
url2 = f"{PIPER_BASE_URL}/{lang_short}/{lang_code}/{fname}"
urllib.request.urlretrieve(url2, fpath)
return onnx_path, json_path
def _piper_synth_chunk(text: str, onnx_path: str, json_path: str, length_scale: float) -> bytes:
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as out_f:
out_path = out_f.name
try:
cmd = [
PIPER_BIN,
"--model", onnx_path,
"--config", json_path,
"--output_file", out_path,
"--length_scale", str(round(length_scale, 2)),
]
result = subprocess.run(
cmd,
input=text.encode("utf-8"),
capture_output=True,
timeout=120,
creationflags=CREATE_NO_WINDOW,
)
if result.returncode != 0:
raise Exception(f"Piper error: {result.stderr.decode()[:200]}")
with open(out_path, "rb") as f:
return f.read()
finally:
if os.path.exists(out_path):
os.unlink(out_path)
def _numpy_concat(parts: list, is_mp3: bool) -> bytes:
if len(parts) == 1:
return parts[0]
import numpy as np
import soundfile as sf
import subprocess, tempfile, os
audio_arrays = []
samplerate = 24000
for data in parts:
fd, in_path = tempfile.mkstemp(suffix=".mp3" if is_mp3 else ".wav")
os.close(fd)
with open(in_path, "wb") as f:
f.write(data)
wav_path = in_path
if is_mp3:
fd, wav_path = tempfile.mkstemp(suffix=".wav")
os.close(fd)
subprocess.run(["ffmpeg", "-y", "-i", in_path, wav_path], check=True, capture_output=True, timeout=60)
try:
arr, sr = sf.read(wav_path)
audio_arrays.append(arr)
samplerate = sr
except Exception:
pass
try: os.unlink(in_path)
except: pass
if is_mp3:
try: os.unlink(wav_path)
except: pass
if not audio_arrays:
return b""
result = audio_arrays[0]
fade_len = int(samplerate * 0.075)
silence_len = int(samplerate * 0.15)
silence = np.zeros(silence_len, dtype=result.dtype)
for next_arr in audio_arrays[1:]:
result = np.concatenate([result, silence])
if len(result) > fade_len and len(next_arr) > fade_len:
fade_out = np.linspace(1.0, 0.0, fade_len)
fade_in = np.linspace(0.0, 1.0, fade_len)
if len(result.shape) > 1:
fade_out = fade_out[:, np.newaxis]
fade_in = fade_in[:, np.newaxis]
overlap_result = result[-fade_len:] * fade_out
overlap_next = next_arr[:fade_len] * fade_in
result[-fade_len:] = overlap_result + overlap_next
result = np.concatenate([result, next_arr[fade_len:]])
else:
result = np.concatenate([result, next_arr])
fd, out_wav = tempfile.mkstemp(suffix=".wav")
os.close(fd)
sf.write(out_wav, result, samplerate)
if is_mp3:
fd, out_mp3 = tempfile.mkstemp(suffix=".mp3")
os.close(fd)
subprocess.run(["ffmpeg", "-y", "-i", out_wav, out_mp3], check=True, capture_output=True, timeout=60)
with open(out_mp3, "rb") as f:
res = f.read()
os.unlink(out_mp3)
os.unlink(out_wav)
return res
else:
with open(out_wav, "rb") as f:
res = f.read()
os.unlink(out_wav)
return res
def _ffmpeg_concat_wav(parts: list) -> bytes:
return _numpy_concat(parts, False)
def synthesize_piper(text: str, voice_code: str, speed: float = 1.0) -> bytes:
if not PIPER_READY:
raise Exception("Piper is not available on this system")
# Only explicitly tested models are permitted.
if voice_code not in PIPER_VOICES:
raise ValueError("Unsupported Piper voice")
onnx_path, json_path = download_piper_model(voice_code)
length_scale = 1.0 / max(0.25, min(4.0, speed))
# Lambi text ko chunk karo (Piper stdin limit + timeout avoid karne ke liye)
if len(text) > 1400:
chunks = split_text(text, max_chars=1400)
else:
chunks = [text]
if len(chunks) == 1:
return _piper_synth_chunk(chunks[0], onnx_path, json_path, length_scale)
parts = []
for ch in chunks:
if ch.strip():
parts.append(_piper_synth_chunk(ch, onnx_path, json_path, length_scale))
if not parts:
raise Exception("Piper: no audio generated")
return _ffmpeg_concat_wav(parts)
def sanitize_text(text: str) -> str:
import re
text = re.sub(r'[\u200B-\u200D\uFEFF]', '', text)
text = re.sub(r'[\x00-\x08\x0b-\x1f\x7f]', '', text)
text = re.sub(r'[ \t]+', ' ', text)
text = re.sub(r'\n\s*\n+', '\n\n', text)
return text.strip()
_EDGE_RATE_RE = re.compile(r'^[+-]?(?:\d+(?:\.\d+)?%|\d+(?:\.\d+)?x)$')
_EDGE_VOLUME_RE = re.compile(r'^[+-]?\d+(?:\.\d+)?%$')
_EDGE_PITCH_RE = re.compile(r'^[+-]?\d+(?:\.\d+)?(?:Hz|st)?$')
def validate_edge_prosody(rate: str, volume: str, pitch: str):
"""Reject anything that could break out of the generated SSML prosody tag."""
if not _EDGE_RATE_RE.match(str(rate or "").strip()):
raise ValueError("Invalid rate. Use a value like +0%, -25% or 1.2x.")
if not _EDGE_VOLUME_RE.match(str(volume or "").strip()):
raise ValueError("Invalid volume. Use a value like +0% or -50%.")
if not _EDGE_PITCH_RE.match(str(pitch or "").strip()):
raise ValueError("Invalid pitch. Use a value like +0Hz or -2st.")
def split_text(text: str, max_chars: int = 1400) -> list:
"""Text ko chunklara bol"""
text = sanitize_text(text)
if not text:
return []
sentence_re = re.compile(
r'(?<=[.!?\u0964\u06D4\u061F\u2026\u3002\uff01\uff1f])\s+(?=\S)'
)
chunks, current = [], ""
for para in re.split(r'\n+', text):
para = para.strip()
if not para:
if current:
chunks.append(current)
current = ""
continue
for sentence in sentence_re.split(para):
sentence = sentence.strip()
if not sentence:
continue
if len(sentence) > max_chars:
words, buf = sentence.split(), ""
for word in words:
if len(word) > max_chars:
if buf:
chunks.append(buf)
buf = ""
for sub_idx in range(0, len(word), max_chars):
chunks.append(word[sub_idx:sub_idx + max_chars])
continue
add = (" " if buf else "") + word
if len(buf) + len(add) <= max_chars:
buf += add
else:
if buf:
chunks.append(buf)
buf = word
if buf:
chunks.append(buf)
elif len(current) + len(sentence) + 1 <= max_chars:
current = (current + " " + sentence).strip()
else:
if current:
chunks.append(current)
current = sentence
if current:
chunks.append(current)
current = ""
if current:
chunks.append(current)
return [c for c in chunks if c.strip()]
# -------------------------------------------------------------------------
# SILERO TTS — v4 model (48kHz, 5 Russian speakers)
# Loaded via torch.package.PackageImporter (requires torch < 2.3)
# -------------------------------------------------------------------------
SILERO_READY = False
SILERO_MODELS_DIR = "/tmp/silero_models"
SILERO_SAMPLE_RATE = 48000
SILERO_MODELS = {}
SILERO_LOAD_LOCK = threading.Lock()
SILERO_MODEL_URLS = [
"https://models.silero.ai/models/tts/ru/v4_ru.pt",
"https://huggingface.co/Derur/silero-models/resolve/main/tts/ru/ru_v4/v4_ru.pt",
]
SILERO_SPEAKERS_RU = [
"aidar", "baya", "kseniya", "xenia", "eugene",
]
def download_silero_model() -> str:
"""Download Silero v4 Russian model. Returns path on success."""
import urllib.request
os.makedirs(SILERO_MODELS_DIR, exist_ok=True)
model_path = os.path.join(SILERO_MODELS_DIR, "v4_ru.pt")
if os.path.exists(model_path) and os.path.getsize(model_path) > 100000:
return model_path
for url in SILERO_MODEL_URLS:
try:
print(f"[DOWNLOAD] Downloading Silero v4: {url[:80]}...")
urllib.request.urlretrieve(url, model_path)
if os.path.getsize(model_path) > 100000:
print(f"[OK] Silero v4 downloaded ({os.path.getsize(model_path)//1024}KB)")
return model_path
os.remove(model_path)
except Exception as e:
print(f"[WARN] Download failed: {e}")
try:
os.remove(model_path)
except Exception:
pass
return ""
def setup_silero():
global SILERO_READY
try:
import torch
model_path = download_silero_model()
if model_path:
model = torch.package.PackageImporter(model_path).load_pickle("tts_models", "model")
SILERO_MODELS["ru"] = model
SILERO_READY = True
print("Silero TTS ready (v4 Russian - 5 speakers)")
else:
print("Silero TTS: model download failed")
except Exception as e:
print(f"[Silero] setup failed (non-critical): {e}")
def synthesize_silero(text: str, voice_code: str) -> bytes:
"""Silero TTS - code: silero:ru_xenia. v4 model via torch.package.
Model lazily load hota hai (self-heal) agar startup thread fail hua ho."""
import numpy as np, scipy.io.wavfile as wav
try:
import torch
except ImportError:
raise Exception("Silero TTS requires torch.")
lang_speaker = voice_code.replace("silero:", "")
lang = lang_speaker.split("_")[0]
speaker = lang_speaker.split("_", 1)[1] if "_" in lang_speaker else lang_speaker
# Validate speaker - only real v4 speakers allowed
valid_speakers = set(SILERO_SPEAKERS_RU)
if speaker not in valid_speakers:
raise Exception(f"Invalid Silero speaker '{speaker}'. Valid: {', '.join(valid_speakers)}")
# Model on-demand load (self-heals if startup thread failed / was slow)
if lang not in SILERO_MODELS:
with SILERO_LOAD_LOCK:
if lang not in SILERO_MODELS:
model_path = download_silero_model()
if not model_path:
raise Exception("Silero v4 model could not be downloaded (check network / model URL).")
model = torch.package.PackageImporter(model_path).load_pickle("tts_models", "model")
SILERO_MODELS[lang] = model
global SILERO_READY
SILERO_READY = True
model = SILERO_MODELS[lang]
audio = model.apply_tts(text=text, speaker=speaker, sample_rate=SILERO_SAMPLE_RATE)
audio_np = audio.numpy() if hasattr(audio, "numpy") else audio.cpu().detach().numpy()
buf = io.BytesIO()
wav.write(buf, SILERO_SAMPLE_RATE, (audio_np * 32767).astype(np.int16))
buf.seek(0)
return buf.read()
def _ffmpeg_concat_mp3(parts: list) -> bytes:
return _numpy_concat(parts, True)
async def _edge_synth_chunk(chunk: str, voice: str, kwargs: dict) -> bytes:
"""One chunk ki audio lao, 3 baar retry karo. Fail par b"" return."""
import edge_tts
import asyncio
for attempt in range(3):
data = bytearray()
try:
comm = edge_tts.Communicate(chunk, voice, **kwargs)
async for packet in comm.stream():
if packet["type"] == "audio" and packet.get("data"):
data.extend(packet["data"])
if data:
return bytes(data)
except Exception:
if attempt == 2:
if voice != "en-US-AriaNeural":
try:
comm = edge_tts.Communicate(chunk, "en-US-AriaNeural", **kwargs)
async for packet in comm.stream():
if packet["type"] == "audio" and packet.get("data"):
data.extend(packet["data"])
if data:
return bytes(data)
except Exception:
pass
return b""
await asyncio.sleep(2)
return b""
async def synthesize_edge(
text: str,
voice: str,
rate: str = "+0%",
volume: str = "+0%",
pitch: str = "+0Hz",
style: str = None,
styledegree: str = None,
) -> bytes:
import edge_tts
chunks = split_text(text)
audio_parts = []
kwargs = {"rate": rate, "volume": volume, "pitch": pitch}
if style and style != "Default" and style != "General":
kwargs["style"] = style
if styledegree is not None:
try:
sd = float(styledegree)
if 0.0 <= sd <= 2.0:
kwargs["styledegree"] = styledegree
except Exception:
pass
# Style drop karne wala safe set (non-EN voices kuch styles reject karti hain)
safe_kwargs = {k: v for k, v in kwargs.items() if k not in ("style", "styledegree")}
for chunk in chunks:
audio = await _edge_synth_chunk(chunk, voice, kwargs)
# Agar style ki wajah se fail hua ho, style hata kar dobara try karo
if not audio and kwargs.get("style"):
audio = await _edge_synth_chunk(chunk, voice, safe_kwargs)
audio_parts.append(audio if audio else b"")
if len(audio_parts) == 1:
return audio_parts[0]
return _ffmpeg_concat_mp3(audio_parts)
# -----------------------------------------------------------------------------
# VOICE CLONING ENGINES - FREE MODELS (Task 4/5)
# -----------------------------------------------------------------------------
CLONE_MODELS = {
"f5tts:v1_base": ("f5tts", "multilingual", "Voice Clone"),
}
_POCKET_MODEL = None
_POCKET_MODEL_ERROR = None
_POCKET_GENERATION_ERROR = None
_POCKET_MODEL_LOCK = threading.Lock()
_POCKET_INFER_LOCK = threading.Lock()
_POCKET_RESULT_CACHE = {}
_POCKET_RESULT_CACHE_ORDER = []
_POCKET_RESULT_CACHE_LOCK = threading.Lock()
_POCKET_RESULT_CACHE_LIMIT = 8
_POCKET_RESULT_CACHE_MAX_BYTES = 16 * 1024 * 1024
_POCKET_VOICE_STATE_CACHE = {}
_POCKET_VOICE_STATE_ORDER = []
_POCKET_VOICE_STATE_LOCK = threading.Lock()
_POCKET_VOICE_STATE_LIMIT = 6
_POCKET_TORCH_PATCHED = False
def _hf_token_configured() -> bool:
return bool(
os.environ.get("HF_TOKEN")
or os.environ.get("HUGGINGFACE_HUB_TOKEN")
or os.environ.get("HUGGING_FACE_HUB_TOKEN")
)
def _pocket_clone_auth_error() -> str:
return (
"Voice cloning weights are gated. Accept the Hugging Face model terms, "
"add a read token as HF_TOKEN in Space secrets, then restart/rebuild the Space."
)
def _pocket_clone_ready() -> bool:
return bool(_POCKET_MODEL is not None and getattr(_POCKET_MODEL, "has_voice_cloning", False))
def _patch_torch_for_pocket_tts():
"""Pocket TTS mutates a streaming KV cache; no_grad keeps tensors mutable."""
global _POCKET_TORCH_PATCHED
if _POCKET_TORCH_PATCHED:
return
try:
import torch
torch.inference_mode = torch.no_grad
_POCKET_TORCH_PATCHED = True
logging.info("Pocket TTS torch compatibility patch enabled")
except Exception as exc:
logging.warning("Pocket TTS torch patch skipped: %s", exc.__class__.__name__)
def _load_pocket_model():
global _POCKET_MODEL, _POCKET_MODEL_ERROR
if _POCKET_MODEL is not None:
return _POCKET_MODEL
with _POCKET_MODEL_LOCK:
if _POCKET_MODEL is not None:
return _POCKET_MODEL
try:
_patch_torch_for_pocket_tts()
from pocket_tts import TTSModel
_POCKET_MODEL = TTSModel.load_model()
# Pocket TTS is inference-only in this service. Disabling gradient
# tracking reduces CPU work and memory without changing synthesis.
try:
import torch
torch.set_grad_enabled(False)
except Exception:
pass
_POCKET_MODEL_ERROR = None
if getattr(_POCKET_MODEL, "has_voice_cloning", False):
print("Pocket TTS CPU voice clone ready")
else:
print("Pocket TTS loaded without voice cloning weights. HF_TOKEN/model access required.")
except Exception as exc:
_POCKET_MODEL_ERROR = str(exc)
logging.exception("Pocket TTS startup failed")
raise RuntimeError(_POCKET_MODEL_ERROR or "Voice clone model is not ready")
return _POCKET_MODEL
def _warm_pocket_model():
if clone_engines_enabled():
try:
_load_pocket_model()
except Exception:
pass
def _reference_audio_key(reference_path: str) -> str:
digest = hashlib.sha256()
with open(reference_path, "rb") as reference_file:
for chunk in iter(lambda: reference_file.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
def _normalize_clone_reference(reference_path: str) -> str:
"""Decode any accepted upload into a mono PCM WAV Pocket TTS can read."""
if not reference_path or not os.path.exists(reference_path):
raise ValueError("A reference audio file is required for voice cloning")
ffmpeg = shutil.which("ffmpeg")
if not ffmpeg:
raise RuntimeError("ffmpeg is required to decode voice clone reference audio")
fd, normalized_path = tempfile.mkstemp(suffix=".wav")
os.close(fd)
try:
result = subprocess.run(
[
ffmpeg,
"-y",
"-hide_banner",
"-loglevel",
"error",
"-i",
reference_path,
"-map",
"0:a:0",
"-t",
"12",
"-vn",
"-sn",
"-dn",
"-acodec",
"pcm_s16le",
"-ac",
"1",
"-ar",
"24000",
"-f",
"wav",
normalized_path,
],
capture_output=True,
timeout=60,
creationflags=CREATE_NO_WINDOW,
)
if result.returncode != 0 or os.path.getsize(normalized_path) <= 44:
detail = result.stderr.decode("utf-8", errors="replace").strip()
raise ValueError(
"Reference audio could not be decoded. "
f"Use a clear WAV, MP3, M4A, OGG, or FLAC file. {detail[:160]}"
)
try:
import soundfile as sf
import numpy as np
data, sr = sf.read(normalized_path)
if len(data.shape) > 1:
data = data.mean(axis=1)
sr = sr or 24000
duration = len(data) / float(sr)
if len(data) == 0 or duration < 0.3:
raise ValueError("Reference audio snippet is too short. Please upload a voice sample of at least 1-3 seconds.")
# Auto-trim if longer than 12s, auto-tile if shorter than 3s
if duration > 12.0:
data = data[:int(12.0 * sr)]
elif duration < 3.0:
repeats = int(np.ceil(3.0 / duration))
data = np.tile(data, repeats)[:int(3.0 * sr)]
rms = np.sqrt(np.mean(data**2))
if rms < 0.0005:
raise ValueError("Reference audio is mostly silence. Please record a clearer voice sample.")
target_rms = 10 ** (-20 / 20)
data = data * (target_rms / (rms + 1e-9))
threshold = 10 ** (-40 / 20)
mask = np.abs(data) > threshold
if np.any(mask):
start = np.argmax(mask)
end = len(mask) - np.argmax(mask[::-1])
data = data[start:end]
sf.write(normalized_path, data, sr)
except ValueError:
raise
except Exception:
pass
return normalized_path
except Exception:
try:
os.unlink(normalized_path)
except OSError:
pass
raise
def _clone_request_key(text: str, reference_key: str) -> str:
payload = f"{reference_key}\0{text or ''}".encode("utf-8")
return hashlib.sha256(payload).hexdigest()
def _get_cached_clone(cache_key: str):
with _POCKET_RESULT_CACHE_LOCK:
audio = _POCKET_RESULT_CACHE.get(cache_key)
if audio is not None:
try:
_POCKET_RESULT_CACHE_ORDER.remove(cache_key)
except ValueError:
pass
_POCKET_RESULT_CACHE_ORDER.append(cache_key)
return audio
def _cache_clone(cache_key: str, audio: bytes):
if len(audio) > _POCKET_RESULT_CACHE_MAX_BYTES:
return
with _POCKET_RESULT_CACHE_LOCK:
if cache_key in _POCKET_RESULT_CACHE:
_POCKET_RESULT_CACHE_ORDER.remove(cache_key)
_POCKET_RESULT_CACHE[cache_key] = audio
_POCKET_RESULT_CACHE_ORDER.append(cache_key)
while len(_POCKET_RESULT_CACHE_ORDER) > _POCKET_RESULT_CACHE_LIMIT:
oldest = _POCKET_RESULT_CACHE_ORDER.pop(0)
_POCKET_RESULT_CACHE.pop(oldest, None)
def _get_cached_voice_state(reference_key: str):
with _POCKET_VOICE_STATE_LOCK:
voice_state = _POCKET_VOICE_STATE_CACHE.get(reference_key)
if voice_state is not None:
try:
_POCKET_VOICE_STATE_ORDER.remove(reference_key)
except ValueError:
pass
_POCKET_VOICE_STATE_ORDER.append(reference_key)
return voice_state
def _cache_voice_state(reference_key: str, voice_state):
with _POCKET_VOICE_STATE_LOCK:
if reference_key in _POCKET_VOICE_STATE_CACHE:
_POCKET_VOICE_STATE_ORDER.remove(reference_key)
_POCKET_VOICE_STATE_CACHE[reference_key] = voice_state
_POCKET_VOICE_STATE_ORDER.append(reference_key)
while len(_POCKET_VOICE_STATE_ORDER) > _POCKET_VOICE_STATE_LIMIT:
oldest = _POCKET_VOICE_STATE_ORDER.pop(0)
_POCKET_VOICE_STATE_CACHE.pop(oldest, None)
def _estimate_min_clone_seconds(text: str) -> float:
words = len(re.findall(r"\S+", str(text or "")))
# A very conservative floor; natural speech can be fast, but a much shorter
# result usually means the model stopped early and skipped words.
return max(0.45, min(20.0, words * 0.105))
def _generate_pocket_chunk(model, voice_state, chunk: str):
import numpy as np
audio = model.generate_audio(voice_state, chunk)
if hasattr(audio, "detach"):
audio = audio.detach().cpu().numpy()
return np.asarray(audio).squeeze().reshape(-1)
def _run_pocket_clone_cpu(text: str, reference_path: str, reference_key: str) -> bytes:
if not reference_path or not os.path.exists(reference_path):
raise ValueError("A reference audio file is required for voice cloning")
model = _load_pocket_model()
if not getattr(model, "has_voice_cloning", False):
raise RuntimeError(_pocket_clone_auth_error())
import numpy as np
voice_state = _get_cached_voice_state(reference_key)
audio_parts = []
with _POCKET_INFER_LOCK:
if voice_state is None:
try:
voice_state = model.get_state_for_audio_prompt(reference_path)
except Exception as exc:
message = str(exc)
if "could not download the weights" in message.lower() or "voice cloning" in message.lower():
raise RuntimeError(_pocket_clone_auth_error()) from exc
raise
_cache_voice_state(reference_key, voice_state)
chunks = split_text(text, max_chars=CLONE_CHUNK_CHARACTERS)
try:
import torch
inference_context = torch.no_grad()
except Exception:
inference_context = nullcontext()
with inference_context:
for chunk in chunks:
audio_np = _generate_pocket_chunk(model, voice_state, chunk)
duration = audio_np.size / float(getattr(model, "sample_rate", 24000) or 24000)
if (
audio_np.size
and len(chunk) > CLONE_RETRY_CHUNK_CHARACTERS
and duration < _estimate_min_clone_seconds(chunk)
):
logging.warning("Clone chunk looked truncated; retrying with smaller chunks")
smaller_parts = []
for sub_chunk in split_text(chunk, max_chars=CLONE_RETRY_CHUNK_CHARACTERS):
sub_audio = _generate_pocket_chunk(model, voice_state, sub_chunk)
if sub_audio.size:
smaller_parts.append(sub_audio)
if smaller_parts:
silence = np.zeros(int(model.sample_rate * 0.08), dtype=smaller_parts[0].dtype)
joined = []
for index, part in enumerate(smaller_parts):
if index:
joined.append(silence)
joined.append(part)
audio_np = np.concatenate(joined)
if audio_np.size:
audio_parts.append(audio_np)
if not audio_parts:
raise RuntimeError("Voice clone model did not produce audio")
if len(audio_parts) == 1:
audio_np = audio_parts[0]
else:
silence = np.zeros(int(model.sample_rate * 0.12), dtype=audio_parts[0].dtype)
joined = []
for index, part in enumerate(audio_parts):
if index:
joined.append(silence)
joined.append(part)
audio_np = np.concatenate(joined)
import scipy.io.wavfile as wav
if audio_np.size == 0:
raise RuntimeError("Voice clone model did not produce audio")
buf = io.BytesIO()
wav.write(buf, int(model.sample_rate), audio_np)
return buf.getvalue()
async def synthesize_f5tts(text: str, reference_path: str = None) -> bytes:
global _POCKET_GENERATION_ERROR
if not reference_path or not os.path.exists(reference_path):
raise ValueError("A reference audio file is required for voice cloning")
try:
reference_key = await asyncio.to_thread(_reference_audio_key, reference_path)
cache_key = _clone_request_key(text, reference_key)
cached = _get_cached_clone(cache_key)
if cached is not None:
return cached
if _get_cached_voice_state(reference_key) is not None:
# Same reference voice is already encoded. Skip ffmpeg normalization for repeated
# desktop batches/generations; audio quality is unchanged because voice_state is reused.
audio = await asyncio.to_thread(
_run_pocket_clone_cpu, text, reference_path, reference_key
)
else:
normalized_path = await asyncio.to_thread(_normalize_clone_reference, reference_path)
try:
audio = await asyncio.to_thread(
_run_pocket_clone_cpu, text, normalized_path, reference_key
)
finally:
try:
os.unlink(normalized_path)
except OSError:
pass
_cache_clone(cache_key, audio)
_POCKET_GENERATION_ERROR = None
return audio
except Exception as exc:
_POCKET_GENERATION_ERROR = exc.__class__.__name__
logging.exception("Voice clone generation failed")
raise RuntimeError("Voice cloning failed on this server. Please contact support.") from exc
def post_process_bytes(audio: bytes, media_type: str) -> bytes:
"""Apply gentle, smooth speech cleanup without block-gate clicks.
A single ffmpeg pass is faster than decoding into Python/SciPy, walking
every 20 ms window, writing a WAV, and encoding again. The conservative
denoise amount removes low-level hiss while preserving voice character.
"""
suffix = ".mp3" if "mpeg" in media_type else ".wav"
fd, in_path = tempfile.mkstemp(suffix=suffix)
os.close(fd)
with open(in_path, "wb") as f:
f.write(audio)
fd, out_path = tempfile.mkstemp(suffix=suffix)
os.close(fd)
try:
ffmpeg = shutil.which("ffmpeg")
if not ffmpeg:
return audio
filters = (
"highpass=f=55,"
"afftdn=nr=4:nf=-55:tn=1,"
"equalizer=f=8000:width_type=h:width=1200:g=1.2,"
"alimiter=limit=0.891:attack=5:release=50"
)
cmd = [
ffmpeg, "-y", "-hide_banner", "-loglevel", "error",
"-i", in_path, "-vn", "-af", filters,
]
if suffix == ".mp3":
# Explicit high-quality encoding avoids ffmpeg's lower default
# bitrate and prevents an avoidable quality drop.
cmd += ["-codec:a", "libmp3lame", "-b:a", "320k", out_path]
else:
cmd += ["-acodec", "pcm_s16le", out_path]
result = subprocess.run(
cmd,
capture_output=True,
timeout=300,
creationflags=CREATE_NO_WINDOW,
)
if result.returncode != 0 or not os.path.exists(out_path) or os.path.getsize(out_path) <= 44:
logging.warning("Audio cleanup failed with ffmpeg exit code %s", result.returncode)
return audio
with open(out_path, "rb") as output_file:
return output_file.read()
except Exception as exc:
logging.warning("Audio cleanup failed: %s", exc.__class__.__name__)
return audio
finally:
for p in (in_path, out_path):
if os.path.exists(p):
try:
os.unlink(p)
except OSError:
pass
@app.post("/tts")
async def tts_endpoint(
request: Request,
engine: str = Form(...), # "edge" | "piper" | "silero" | "f5tts"
text: str = Form(...),
voice: str = Form("en-US-AvaNeural"), # edge voice code OR piper/silero code
rate: str = Form("+0%"), # edge only
volume: str = Form("+0%"), # edge only
pitch: str = Form("+0Hz"), # edge only
speed: float = Form(1.0), # piper only
style: str = Form(None), # edge style (emotion)
styledegree: str = Form(None), # edge style degree 0-2
post_process: bool = Form(True), # audio post-processing toggle
voice_cloning: bool = Form(False), # voice cloning toggle
reference_audio: UploadFile = File(None),
):
if not text or not text.strip():
return JSONResponse(status_code=400, content={"error": "Text is empty"})
text = text.strip()
# Abuse / timeout guard - lambi text Chapter Mode se bhejo
if len(text) > 60000:
return JSONResponse(
status_code=413,
content={"error": "Text too long (max 60000 chars). Use Chapter Mode for longer text."},
)
engine = engine.strip().lower()
if engine not in available_engines():
return JSONResponse(status_code=400, content={"error": "Unsupported TTS engine"})
has_reference = bool(reference_audio is not None and reference_audio.filename)
if has_reference and engine not in CLONE_ENGINES:
return JSONResponse(
status_code=400,
content={"error": "Reference audio is only accepted for voice cloning"},
)
if engine in CLONE_ENGINES and not has_reference:
return JSONResponse(
status_code=400,
content={"error": "Voice cloning requires reference audio"},
)
if engine == "piper" and voice not in PIPER_VOICES:
return JSONResponse(status_code=400, content={"error": "Unsupported Piper voice"})
if engine == "silero" and not str(voice).startswith("silero:ru_"):
return JSONResponse(status_code=400, content={"error": "Unsupported Silero voice"})
if engine == "edge" and not re.match(r"^[a-z]{2,3}-[A-Z]{2}-.+Neural$", str(voice)):
return JSONResponse(status_code=400, content={"error": "Unsupported Edge voice"})
if engine == "edge":
try:
validate_edge_prosody(rate, volume, pitch)
except ValueError as exc:
return JSONResponse(status_code=400, content={"error": str(exc)})
# These two checks run BEFORE authorize_request on purpose. Authorization
# books the requested characters against the monthly allowance, and only
# the try/finally further down gives them back. Rejecting here afterwards
# would leak the reservation and silently bill the customer for a request
# that never ran.
if engine in CLONE_ENGINES and not clone_backend_available():
return JSONResponse(
status_code=403,
content={"error": "Voice cloning is not available on this server"},
)
if engine in CLONE_ENGINES and len(text) > MAX_CLONE_CHARACTERS:
return JSONResponse(
status_code=413,
content={
"error": (
"Clone text is too long "
f"(max {MAX_CLONE_CHARACTERS} characters per request)"
)
},
)
auth_context = await authorize_request(request, engine, len(text))
reference_path = None
started_at = time.monotonic()
generation_status = "error"
slot_acquired = False
try:
acquire_worker_slot()
slot_acquired = True
if has_reference:
suffix = os.path.splitext(reference_audio.filename)[1].lower() or ".wav"
if suffix not in (".wav", ".mp3", ".m4a", ".ogg", ".flac"):
return JSONResponse(status_code=400, content={"error": "Unsupported reference audio format"})
fd_ref, reference_path = tempfile.mkstemp(suffix=suffix)
os.close(fd_ref)
data = await reference_audio.read(25 * 1024 * 1024 + 1)
if not data:
return JSONResponse(status_code=400, content={"error": "Reference audio is empty"})
if len(data) > 25 * 1024 * 1024:
return JSONResponse(status_code=413, content={"error": "Reference audio is too large (max 25 MB)"})
with open(reference_path, "wb") as f:
f.write(data)
voice_cloning = True
if engine == "edge":
audio = await synthesize_edge(text, voice, rate=rate, volume=volume, pitch=pitch, style=style, styledegree=styledegree)
media = "audio/mpeg"
fname = "tts_edge.mp3"
elif engine == "piper":
# Sync + subprocess/torch -> run off the event loop so the server
# stays responsive and doesn't time out under load.
audio = await asyncio.to_thread(synthesize_piper, text, voice, speed)
media = "audio/wav"
fname = "tts_piper.wav"
elif engine == "silero":
audio = await asyncio.to_thread(synthesize_silero, text, voice)
media = "audio/wav"
fname = "tts_silero.wav"
elif engine == "f5tts":
audio = await synthesize_f5tts(text, reference_path=reference_path)
media = "audio/wav"
fname = "tts_f5_clone.wav"
else:
return JSONResponse(status_code=400, content={"error": f"Unknown engine: {engine}"})
# Preserve clone identity and improve speed: Pocket TTS already emits
# clean PCM. Re-filtering cloned audio can subtly change timbre and
# adds a full ffmpeg pass. Normal TTS still receives gentle cleanup.
if post_process and engine not in CLONE_ENGINES:
audio = await asyncio.to_thread(post_process_bytes, audio, media)
if not audio or len(audio) == 0:
return JSONResponse(
status_code=502,
content={"error": "Audio generation failed or produced empty audio. Please retry."},
)
generation_status = "success"
return StreamingResponse(
io.BytesIO(audio),
media_type=media,
headers={"Content-Disposition": f"attachment; filename={fname}"},
)
except HTTPException:
# "Worker is busy; try another Space" (503) is raised by
# acquire_worker_slot. Swallowing it into a generic 500 stopped the
# desktop client from failing over to the next Space in the pool.
raise
except Exception as e:
logging.error(f"TTS error: {e}", exc_info=True)
return JSONResponse(
status_code=500,
content={"error": "Audio generation failed. Please retry shortly."},
)
finally:
await asyncio.to_thread(
record_usage,
auth_context,
engine,
len(text),
generation_status,
int((time.monotonic() - started_at) * 1000),
)
if slot_acquired:
release_worker_slot()
try:
if reference_path and os.path.exists(reference_path):
os.unlink(reference_path)
except Exception:
pass
from fastapi import Body
@app.post("/register_user")
async def register_user(request: Request, body: dict = Body(...)):
db = _require_firebase_db()
auth_header = request.headers.get("Authorization", "")
if not auth_header.startswith("Bearer "):
raise HTTPException(401, "Missing authentication token")
token = auth_header.split("Bearer ", 1)[1]
try:
decoded = fb_auth.verify_id_token(token)
except Exception:
raise HTTPException(401, "Invalid token")
uid = decoded["uid"]
user_ref = db.collection("users").document(uid)
user_doc = user_ref.get()
data = {
"email": decoded.get("email", ""),
"display_name": body.get("display_name", ""),
"updated_at": firestore.SERVER_TIMESTAMP
}
device_id = str(body.get("device_id", "")).strip()[:128]
if not user_doc.exists:
data["created_at"] = firestore.SERVER_TIMESTAMP
data["is_blocked"] = False
data["device_ids"] = [device_id] if device_id else []
user_ref.set(data)
else:
existing = user_doc.to_dict() or {}
license_key = existing.get("license_key")
if device_id:
if not license_key:
# Keep only the current pre-activation device. Otherwise an
# unlicensed account can accumulate stale IDs and later fail
# activation against a one-device license.
data["device_ids"] = [device_id]
else:
license_doc = db.collection("licenses").document(license_key).get()
lic = license_doc.to_dict() if license_doc.exists else {}
max_devices = max(1, int(lic.get("max_devices") or 1))
_register_device_tx(db, user_ref, device_id, max_devices)
user_ref.update(data)
return {"status": "success", "uid": uid}
@app.post("/activate_license")
async def activate_license(request: Request, body: dict = Body(...)):
db = _require_firebase_db()
auth_header = request.headers.get("Authorization", "")
if not auth_header.startswith("Bearer "):
raise HTTPException(401, "Missing token")
token = auth_header.split("Bearer ", 1)[1]
try:
decoded = fb_auth.verify_id_token(token)
except Exception:
raise HTTPException(401, "Invalid token")
uid = decoded["uid"]
license_key = str(body.get("license_key") or "").strip().upper()
if not license_key:
raise HTTPException(400, "license_key is required")
if not re.fullmatch(r"[A-Z0-9][A-Z0-9_-]{5,79}", license_key):
raise HTTPException(400, "License key format is invalid")
license_ref = db.collection("licenses").document(license_key)
license_doc = license_ref.get()
if not license_doc.exists:
raise HTTPException(404, "License key not found")
lic = license_doc.to_dict()
current_status = lic.get("status")
if current_status not in ("unused", "unclaimed", "active"):
raise HTTPException(400, f"License status is {current_status}")
if lic.get("assigned_uid") and lic.get("assigned_uid") != uid:
raise HTTPException(400, "License is already assigned to another user")
from datetime import datetime, timedelta, timezone
duration_days = lic.get("duration_days", 30)
expiry_date = lic.get("expiry_date")
device_id = str(request.headers.get("X-Device-ID", "")).strip()[:128]
user_ref = db.collection("users").document(uid)
@firestore.transactional
def claim_license(transaction):
fresh_snapshot = license_ref.get(transaction=transaction)
if not fresh_snapshot.exists:
raise HTTPException(404, "License key not found")
fresh = fresh_snapshot.to_dict() or {}
fresh_status = fresh.get("status")
if fresh_status not in ("unused", "unclaimed", "active"):
raise HTTPException(400, f"License status is {fresh_status}")
assigned_uid = fresh.get("assigned_uid")
if assigned_uid and assigned_uid != uid:
raise HTTPException(400, "License is already assigned to another user")
raw_claimed_expiry = fresh.get("expiry_date")
updates = {
"status": "active",
"assigned_uid": uid,
"assigned_email": decoded.get("email", ""),
}
if raw_claimed_expiry in (None, ""):
claimed_expiry = datetime.now(timezone.utc) + timedelta(
days=int(fresh.get("duration_days") or duration_days)
)
updates["activated_at"] = firestore.SERVER_TIMESTAMP
updates["expiry_date"] = claimed_expiry
else:
claimed_expiry = _coerce_utc_datetime(raw_claimed_expiry)
if claimed_expiry is None:
raise HTTPException(400, "License expiry is invalid. Please contact support")
if claimed_expiry < datetime.now(timezone.utc):
raise HTTPException(400, "License has expired")
# Normalize imported ISO strings back to a real Firestore timestamp.
if isinstance(raw_claimed_expiry, str):
updates["expiry_date"] = claimed_expiry
fresh_user_snapshot = user_ref.get(transaction=transaction)
fresh_user = fresh_user_snapshot.to_dict() if fresh_user_snapshot.exists else {}
device_ids = list(fresh_user.get("device_ids") or [])
max_devices = max(1, int(fresh.get("max_devices") or 1))
if len(device_ids) > max_devices:
raise HTTPException(403, "Maximum licensed devices reached")
if device_id and device_id not in device_ids:
if len(device_ids) >= max_devices:
raise HTTPException(403, "Maximum licensed devices reached")
device_ids.append(device_id)
user_clone_setting = (
bool(fresh_user.get("voice_clone_enabled"))
if "voice_clone_enabled" in fresh_user
else bool(fresh.get("voice_clone", False))
)
transaction.update(license_ref, updates)
transaction.set(user_ref, {
"license_key": license_key,
"license_status": "active",
"license_expiry": claimed_expiry,
"voice_clone_enabled": user_clone_setting,
"device_ids": device_ids,
"updated_at": firestore.SERVER_TIMESTAMP,
}, merge=True)
return claimed_expiry, user_clone_setting, fresh, dict(fresh_user, device_ids=device_ids)
expiry_date, user_clone_setting, lic, user_data = claim_license(db.transaction())
# User-specific pool: dedicated Spaces first, otherwise public auto pool.
space_pool = _license_space_pool(db, uid, user_data, license_key, lic)
credit_summary = _monthly_credit_summary(db, uid, lic)
return {
"status": "success",
"plan_name": lic.get("plan_name", "Standard"),
"expiry_date": str(expiry_date) if expiry_date else None,
"voice_clone_enabled": user_clone_setting and bool(lic.get("voice_clone", False)),
"app_link": lic.get("app_link") or lic.get("appLink", ""),
"space_pool": space_pool,
"space_pool_exclusive": _space_pool_is_exclusive(user_data, lic),
**credit_summary,
}
@app.get("/license_status")
async def license_status(request: Request):
db = _require_firebase_db()
auth_header = request.headers.get("Authorization", "")
if not auth_header.startswith("Bearer "):
raise HTTPException(401, "Missing token")
token = auth_header.split("Bearer ", 1)[1]
try:
decoded = fb_auth.verify_id_token(token)
except Exception:
raise HTTPException(401, "Invalid token")
uid = decoded["uid"]
user_doc = db.collection("users").document(uid).get()
if not user_doc.exists:
return {"has_license": False, "status": "none"}
user_data = user_doc.to_dict()
if user_data.get("is_blocked"):
return {"has_license": False, "status": "blocked"}
license_key = user_data.get("license_key")
if not license_key:
return {"has_license": False, "status": "none"}
license_doc = db.collection("licenses").document(license_key).get()
if not license_doc.exists:
return {"has_license": False, "status": "none"}
lic = license_doc.to_dict()
lic_status = lic.get("status", "active")
if lic_status != "active":
return {"has_license": False, "status": lic_status}
from datetime import datetime, timezone
raw_expiry_date = lic.get("expiry_date")
expiry_date = _coerce_utc_datetime(raw_expiry_date)
if raw_expiry_date not in (None, "") and expiry_date is None:
return {"has_license": False, "status": "invalid_expiry", "is_expired": False}
days_remaining = 30
if expiry_date:
delta = expiry_date - datetime.now(timezone.utc)
days_remaining = delta.days
if delta.total_seconds() < 0:
return {"has_license": False, "status": "expired", "is_expired": True}
# User-specific pool: dedicated Spaces first, otherwise public auto pool.
space_pool = _license_space_pool(db, uid, user_data, license_key, lic)
credit_summary = _monthly_credit_summary(db, uid, lic)
return {
"has_license": True,
"status": "active",
"license_key": license_key,
"plan_name": lic.get("plan_name", "Pro"),
"expiry_date": str(expiry_date) if expiry_date else None,
"days_remaining": max(0, days_remaining),
"voice_clone_enabled": bool(user_data.get("voice_clone_enabled", lic.get("voice_clone", False))) and bool(lic.get("voice_clone", False)),
"app_link": lic.get("app_link") or lic.get("appLink", ""),
"space_pool": space_pool,
"space_pool_exclusive": _space_pool_is_exclusive(user_data, lic),
"is_expired": False,
**credit_summary,
}
@app.get("/status")
def status():
return {"status": "ok"}
@app.get("/health")
@app.head("/health")
def health():
with _SPACE_STATE_LOCK:
active_jobs = _ACTIVE_JOBS
policy = get_runtime_policy()
clone_status = clone_backend_status()
firebase_configured = bool(HAS_FIREBASE and _init_firebase())
if AUTH_MODE == "api_secret":
auth_ready = bool(API_SECRET)
elif AUTH_MODE == "layered":
auth_ready = firebase_configured and bool(API_SECRET)
else:
auth_ready = firebase_configured
worker_available = (
auth_ready
and active_jobs < MAX_CONCURRENT_JOBS
and bool(policy.get("space_enabled", True))
and not bool(policy.get("maintenance_mode", False))
)
return {
"status": "ok" if auth_ready else "degraded",
"space_id": SPACE_ID,
"firebase_configured": firebase_configured,
"auth_ready": auth_ready,
"active_jobs": active_jobs,
"max_concurrent_jobs": MAX_CONCURRENT_JOBS,
"available": worker_available,
"clone_enabled": bool(policy.get("space_clone_enabled", False)),
"clone_ready": bool(clone_status.get("ready", False)),
"clone_status": clone_status.get("message", "unknown"),
"engines": BASE_ENGINES + (CLONE_ENGINES if policy.get("space_clone_enabled", False) else []),
"maintenance_mode": bool(policy.get("maintenance_mode", False)),
"version": SERVICE_VERSION,
}
@app.get("/all_voices")
async def all_voices_list(request: Request):
"""All configured Edge, Piper, Silero, and CPU voice-cloning voices."""
await authorize_request(request, "edge")
result = {"edge": {}, "piper": {}, "silero": {}}
# Current Edge catalog. Desktop defaults to a smaller Featured view.
try:
import edge_tts
voices = await edge_tts.list_voices()
for v in voices:
short = v.get("ShortName", "")
friendly = v.get("FriendlyName", "") or short
name = friendly
for remove in ["Microsoft Server Speech Text to Speech Voice", "Microsoft", "Online", "(Natural)", "(Neural)", "(Standard)", "(Multilingual)", "(Expressive)"]:
name = name.replace(remove, "")
if "," in name:
name = name.split(",")[-1]
# Clean dash pattern: " - " or " - " -> single " - "
name = re.sub(r'\s*-\s*', ' - ', name)
# Collapse all whitespace to single space
name = re.sub(r'\s+', ' ', name)
name = name.strip(" -").strip()
region = short.split("-")[0] + "-" + short.split("-")[1] if "-" in short else ""
result["edge"][f"{name} [{region}]"] = short
except Exception as e:
print(f"Edge voices error: {e}")
# Curated Piper catalog only; arbitrary repository models are not exposed.
piper_display_names = {
"piper:en_US-amy-medium": "Amy [en-US]",
"piper:en_US-joe-medium": "Joe [en-US]",
"piper:en_US-lessac-medium": "Lessac HQ [en-US]",
"piper:en_US-ryan-high": "Ryan HQ [en-US]",
"piper:en_GB-alan-medium": "Alan [en-GB]",
"piper:en_GB-alba-medium": "Alba [en-GB]",
"piper:ur_PK-fasih-medium": "Fasih [ur-PK]",
"piper:ar_JO-kareem-medium": "Kareem [ar-JO]",
"piper:hi_IN-pratham-medium": "Pratham [hi-IN]",
"piper:de_DE-thorsten-medium": "Thorsten [de-DE]",
"piper:ru_RU-irina-medium": "Irina [ru-RU]",
"piper:fr_FR-upmc-medium": "UPMC HQ [fr-FR]",
"piper:pt_BR-faber-medium": "Faber [pt-BR]",
"piper:tr_TR-dfki-medium": "Dfki [tr-TR]",
"piper:nl_NL-mls-medium": "MLS [nl-NL]",
}
for code in PIPER_VOICES:
result["piper"][piper_display_names[code]] = code
# Silero v4 - Russian (official v4_ru speakers)
silero_ru_speakers = {
"aidar": "Aidar", "baya": "Baya", "kseniya": "Kseniya",
"xenia": "Xenia", "eugene": "Eugene",
}
for speaker_code, display_name in silero_ru_speakers.items():
result["silero"][f"{display_name} \u2022 Russian [RU]"] = f"silero:ru_{speaker_code}"
if not clone_backend_available():
total = sum(len(group) for group in result.values())
return {"voices": result, "total": total, "clone_enabled": False}
result["f5tts"] = {"Voice Clone": "f5tts:v1_base"}
total = sum(len(group) for group in result.values())
return {"voices": result, "total": total, "clone_enabled": True}
@app.get("/")
def root():
return {"name": "VoiceCraft Service", "status": "ok"}
if __name__ == "__main__":
import uvicorn
port = int(os.environ.get("PORT") or os.environ.get("GRADIO_SERVER_PORT") or 7860)
uvicorn.run(app, host="0.0.0.0", port=port, log_level="info")