voxi-lite / server.py
Joshuant's picture
Voxi Lite 0.5.0: varun, vijay, vysh; laughs and fillers; sentence endings; Indian-English fixes; API key and /ui
5f6a457 verified
Raw History Blame Contribute Delete
12.5 kB
"""Voxi Lite HTTP / WebSocket server.
uvicorn server:app --host 0.0.0.0 --port 8080 (one model per process; scale with more processes)
Environment:
VOXI_MODEL_DIR model folder (default: this folder)
VOXI_THREADS CPU threads per process (default 4)
VOXI_WARM JSON file: {"hi": ["नमस्ते, ..."], "en": ["Hello, ..."]} phrases pre-rendered at start-up
VOXI_WARM_VOICES comma list of voices to warm (default payal,arjun)
VOXI_API_KEY if set, every endpoint except /health needs it: header "Authorization: Bearer KEY" or
"X-API-Key: KEY" (WebSocket: either header, or ?key=KEY for browsers); otherwise 401
Endpoints:
GET /health {"status": "ok", "version", "voices"}
GET /ui browser test console (ui.html; no key needed to load it, the page asks for it)
GET /v1/voices voice names (base voices and their styles; harper, ethan, amelia, oliver: lang "en" only)
POST /v1/tts {"text", "lang": "hi" (hi ta en hinglish tanglish), "voice": "payal", "format": "wav|pcm16|mulaw",
"sample_rate": 8000|16000|24000, "stream": false}
stream=true: chunked raw audio (pcm16 / mulaw) as it is made
POST /v1/audio/speech OpenAI-style: {"model", "input", "voice", "response_format": "wav|pcm",
"language": "hi"} (pcm = 24 kHz 16-bit little-endian, as OpenAI)
WS /v1/tts/ws send {"text", "lang", "voice", "format", "sample_rate", "request_id"};
receive binary audio frames, then {"event": "done", "request_id", ...}.
Send {"event": "cancel"} to stop the current request. Several requests per socket.
Errors: 400 with {"error": "..."} for bad input (unknown voice / language / format, text too long).
Logs: one JSON line per request on the "voxi" logger with lengths, language, voice, timings and status; never the
text or the audio.
"""
import asyncio
import contextlib
import hmac
import json
import logging
import os
import threading
import time
from pathlib import Path
from fastapi import FastAPI, Request, WebSocket, WebSocketDisconnect
from fastapi.responses import HTMLResponse, JSONResponse, Response, StreamingResponse
from pydantic import BaseModel
from voxi import FORMATS, MAX_CHARS, RATES, SR, VoxiLite, encode
HERE = Path(__file__).resolve().parent
log = logging.getLogger("voxi")
if not log.handlers:
h = logging.StreamHandler(); h.setFormatter(logging.Formatter("%(message)s")); log.addHandler(h); log.setLevel(logging.INFO)
MIME = {"wav": "audio/wav", "pcm16": "audio/L16", "mulaw": "audio/basic"}
class TTSRequest(BaseModel):
text: str
lang: str = "hi"
voice: str = "payal"
format: str = "wav"
sample_rate: int = SR
stream: bool = False
class SpeechRequest(BaseModel): # OpenAI /v1/audio/speech shape
input: str
voice: str = "payal"
model: str = "voxi-lite"
response_format: str = "wav"
language: str = "hi"
speed: float = 1.0
class Engine:
"""One VoxiLite per process; synthesis is serialised (the model is not thread-safe), chunk by chunk, so
concurrent streams interleave instead of waiting for each other's whole reply."""
def __init__(self):
self.tts = VoxiLite(os.environ.get("VOXI_MODEL_DIR", HERE), threads=int(os.environ.get("VOXI_THREADS", 4)))
self.lock = threading.Lock()
self.version = self.tts.tts.cfg.get("version", "?")
with self.lock:
self.tts.synthesize("नमस्ते।", "hi") # warm-up
warm = os.environ.get("VOXI_WARM")
if warm:
voices = os.environ.get("VOXI_WARM_VOICES", "payal,arjun").split(",")
for lang, phrases in json.load(open(warm, encoding="utf-8")).items():
with self.lock:
self.tts.warm(phrases, lang, voices)
def check(self, text, lang, voice, fmt, rate):
if fmt not in FORMATS:
raise ValueError(f"unsupported format {fmt!r}; use one of {list(FORMATS)}")
if rate not in RATES:
raise ValueError(f"unsupported sample_rate {rate}; use one of {list(RATES)}")
self.tts._check_voice(voice, lang)
self.tts.prepare(text, lang) # raises ValueError for a bad language or long text
def whole(self, text, lang, voice):
with self.lock:
return self.tts.synthesize(text, lang, voice)
def chunks(self, text, lang, voice):
"""Thread-side generator: each chunk is made under the lock."""
it = self.tts.stream(text, lang, voice)
while True:
with self.lock:
y = next(it, None)
if y is None:
return
yield y
engine: Engine | None = None
@contextlib.asynccontextmanager
async def lifespan(_app):
global engine
if engine is None:
engine = await asyncio.to_thread(Engine)
yield
app = FastAPI(title="Voxi Lite", version="0.4.0", lifespan=lifespan)
API_KEY = os.environ.get("VOXI_API_KEY", "")
def _key_ok(headers, query=None):
if not API_KEY:
return True
auth = headers.get("authorization", "")
got = auth[7:] if auth.lower().startswith("bearer ") else headers.get("x-api-key", "") or (query or {}).get("key", "")
return hmac.compare_digest(got.encode(), API_KEY.encode())
@app.middleware("http")
async def require_key(request: Request, call_next):
if request.url.path not in ("/health", "/ui") and not _key_ok(request.headers):
return JSONResponse({"error": "missing or wrong API key"}, status_code=401)
return await call_next(request)
def _log(**kw):
log.info(json.dumps({"event": "tts", **kw}, ensure_ascii=False))
def _bad(e):
return JSONResponse({"error": str(e)}, status_code=400)
@app.get("/health")
def health():
return {"status": "ok", "version": engine.version, "voices": sum("@" not in v for v in engine.tts.voices)}
@app.get("/ui", response_class=HTMLResponse)
def ui():
page = HERE / "ui.html"
return HTMLResponse(page.read_text(encoding="utf-8") if page.exists() else "<p>ui.html not installed</p>")
@app.get("/v1/voices")
def voices():
return {"voices": engine.tts.voices}
async def _stream_bytes(text, lang, voice, fmt, rate, meta):
t0, first, n = time.perf_counter(), None, 0
gen = engine.chunks(text, lang, voice)
try:
while True:
y = await asyncio.to_thread(next, gen, None)
if y is None:
break
first = first if first is not None else time.perf_counter() - t0
n += len(y)
yield encode(y, "pcm16" if fmt == "wav" else fmt, rate)
meta.update(status="ok")
except Exception as e: # noqa: BLE001
meta.update(status="error", error=type(e).__name__)
raise
finally:
_log(**meta, first_ms=None if first is None else round(first * 1000), total_ms=round((time.perf_counter() - t0) * 1000),
audio_s=round(n / SR, 2))
@app.post("/v1/tts")
async def tts(req: TTSRequest):
meta = {"api": "tts", "chars": len(req.text), "lang": req.lang, "voice": req.voice, "format": req.format,
"sample_rate": req.sample_rate, "stream": req.stream}
try:
engine.check(req.text, req.lang, req.voice, req.format, req.sample_rate)
if req.stream and req.format == "wav":
raise ValueError("stream=true needs format pcm16 or mulaw (a WAV header needs the total length)")
except ValueError as e:
_log(**meta, status="bad_request", error=str(e)[:120]); return _bad(e)
if req.stream:
mime = f"{MIME[req.format]};rate={req.sample_rate}"
return StreamingResponse(_stream_bytes(req.text, req.lang, req.voice, req.format, req.sample_rate, meta), media_type=mime)
t0 = time.perf_counter()
y = await asyncio.to_thread(engine.whole, req.text, req.lang, req.voice)
_log(**meta, status="ok", total_ms=round((time.perf_counter() - t0) * 1000), audio_s=round(len(y) / SR, 2))
mime = MIME[req.format] + ("" if req.format == "wav" else f";rate={req.sample_rate}")
return Response(encode(y, req.format, req.sample_rate), media_type=mime)
@app.post("/v1/audio/speech")
async def speech(req: SpeechRequest):
fmt = {"wav": "wav", "pcm": "pcm16"}.get(req.response_format)
meta = {"api": "speech", "chars": len(req.input), "lang": req.language, "voice": req.voice, "format": req.response_format}
try:
if fmt is None:
raise ValueError("response_format must be wav or pcm")
if req.speed != 1.0:
raise ValueError("speed other than 1.0 is not supported")
engine.check(req.input, req.language, req.voice, fmt, SR)
except ValueError as e:
_log(**meta, status="bad_request", error=str(e)[:120]); return _bad(e)
t0 = time.perf_counter()
y = await asyncio.to_thread(engine.whole, req.input, req.language, req.voice)
_log(**meta, status="ok", total_ms=round((time.perf_counter() - t0) * 1000), audio_s=round(len(y) / SR, 2))
return Response(encode(y, fmt, SR), media_type=MIME[fmt] + ("" if fmt == "wav" else f";rate={SR}"))
@app.websocket("/v1/tts/ws")
async def tts_ws(ws: WebSocket):
if not _key_ok(ws.headers, ws.query_params):
await ws.close(code=1008); return # policy violation: missing or wrong API key
await ws.accept()
cancel = asyncio.Event()
queue: asyncio.Queue = asyncio.Queue()
async def reader(): # requests and cancels arrive while audio is being sent
try:
while True:
msg = json.loads(await ws.receive_text())
if msg.get("event") == "cancel":
cancel.set()
else:
await queue.put(msg)
except (WebSocketDisconnect, RuntimeError):
cancel.set(); await queue.put(None)
except json.JSONDecodeError:
await queue.put({"_bad": "messages must be JSON"})
task = asyncio.create_task(reader())
try:
while True:
msg = await queue.get()
if msg is None:
break
rid = msg.get("request_id")
if "_bad" in msg:
await ws.send_text(json.dumps({"event": "error", "request_id": rid, "error": msg["_bad"]})); continue
text, lang, voice = msg.get("text", ""), msg.get("lang", "hi"), msg.get("voice", "payal")
fmt, rate = msg.get("format", "pcm16"), int(msg.get("sample_rate", SR))
meta = {"api": "ws", "chars": len(text), "lang": lang, "voice": voice, "format": fmt, "sample_rate": rate}
try:
if fmt == "wav":
raise ValueError("use format pcm16 or mulaw on the WebSocket")
if len(text) > MAX_CHARS:
raise ValueError(f"text longer than {MAX_CHARS} characters")
engine.check(text, lang, voice, fmt, rate)
except ValueError as e:
_log(**meta, status="bad_request", error=str(e)[:120])
await ws.send_text(json.dumps({"event": "error", "request_id": rid, "error": str(e)})); continue
cancel.clear()
t0, first, n, status = time.perf_counter(), None, 0, "ok"
gen = engine.chunks(text, lang, voice)
while True:
if cancel.is_set():
status = "cancelled"; break
y = await asyncio.to_thread(next, gen, None)
if y is None:
break
first = first if first is not None else time.perf_counter() - t0
n += len(y)
await ws.send_bytes(encode(y, fmt, rate))
done = {"event": "done", "request_id": rid, "status": status, "audio_s": round(n / SR, 2),
"first_audio_ms": None if first is None else round(first * 1000)}
_log(**meta, status=status, first_ms=done["first_audio_ms"], total_ms=round((time.perf_counter() - t0) * 1000),
audio_s=done["audio_s"])
await ws.send_text(json.dumps(done))
except (WebSocketDisconnect, RuntimeError):
pass
finally:
task.cancel()