app.py

import asyncio
import json
import os
import sys
import shutil
import time
from pathlib import Path
from typing import Dict, Optional, Tuple

import uvicorn
from fastapi import FastAPI, WebSocket, WebSocketDisconnect, Request, BackgroundTasks, Query, Header
from fastapi.responses import JSONResponse, FileResponse, PlainTextResponse

# -------------------------
# Writable directories
# -------------------------

def pick_writable_dir(candidates):
    errs = []
    for p in candidates:
        if not p:
            continue
        try:
            p.mkdir(parents=True, exist_ok=True)
            probe = p / ".probe"
            probe.write_bytes(b"ok")
            probe.unlink(missing_ok=True)
            return p
        except Exception as e:
            errs.append(f"{p}: {type(e).__name__}({e})")
    raise RuntimeError("No writable dir. Tried:\n  " + "\n  ".join(errs))

ENV_DIR = os.getenv("TTS_DATA_DIR")
VOICE_CANDIDATES = [
    Path("/tmp/actualtts/voices"),
    Path("/dev/shm/actualtts_voices"),
    Path(ENV_DIR) / "voices" if ENV_DIR else None,
]
FILE_CANDIDATES = [
    Path("/tmp/actualtts/files"),
    Path("/dev/shm/actualtts_files"),
    Path(ENV_DIR) / "files" if ENV_DIR else None,
]

VOICES_DIR = pick_writable_dir([p for p in VOICE_CANDIDATES if p])
FILES_DIR  = pick_writable_dir([p for p in FILE_CANDIDATES if p])

def _safe_unlink(path: Path):
    try:
        path.unlink(missing_ok=True)
    except Exception:
        pass

# -------------------------
# Piper command resolution
# -------------------------

def resolve_piper_cmd():
    env = os.getenv("PIPER_BIN")
    if env:
        return env.split()
    path = shutil.which("piper")
    if path:
        return [path]
    # fallback to module runner if console script isn't on PATH
    return [sys.executable, "-m", "piper"]

PIPER_CMD = resolve_piper_cmd()

# -------------------------
# Streaming / synthesis tuning
# -------------------------

# Intra-sentence silence added by Piper between sentences (seconds)
SENTENCE_SILENCE = float(os.getenv("PIPER_SENTENCE_SILENCE", "0.05"))  # 50 ms

# Server-to-client WS frame size target in milliseconds (steady pacing)
STREAM_BATCH_MS  = int(os.getenv("STREAM_BATCH_MS", "100"))            # ~100 ms frames

DEFAULT_CH = 1  # mono

# NEW: Prebuffer to avoid pauses (synthesize immediately, start streaming later)
PREBUFFER_MS = int(os.getenv("PREBUFFER_MS", "6000"))                  # ~6s buffer before first bytes
PREBUFFER_MAX_WAIT_MS = int(os.getenv("PREBUFFER_MAX_WAIT_MS", "15000"))  # fail-safe cap

# Input clamp (basic DoS protection)
MAX_TEXT_CHARS = int(os.getenv("MAX_TEXT_CHARS", "800"))

# Optional shared secret (x-auth header) for internal/protected calls
AUTH_SHARED_SECRET = (os.getenv("AUTH_SHARED_SECRET") or "").strip()

def _auth_ok(x_auth: Optional[str]) -> bool:
    return (not AUTH_SHARED_SECRET) or (x_auth == AUTH_SHARED_SECRET)

# -------------------------
# Voice download & checks
# -------------------------

HF_REPO_BASE = "https://huggingface.co/rhasspy/piper-voices/resolve"
HF_REV       = os.getenv("PIPER_VOICES_REV", "main")  # optionally pin a commit hash

# sanity thresholds (bytes)
MIN_ONNX_BYTES = int(os.getenv("MIN_ONNX_BYTES", "5000000"))  # >= ~5MB
MIN_JSON_BYTES = int(os.getenv("MIN_JSON_BYTES", "1000"))     # >= 1KB

# (lang, country, family, quality, basename)
VOICE_MAP: Dict[str, Tuple[str, str, str, str, str]] = {
    "en_US-libritts-high": ("en", "en_US", "libritts", "high",   "en_US-libritts-high"),
    "en_US-lessac-high":   ("en", "en_US", "lessac",   "high",   "en_US-lessac-high"),
    "en_US-amy-medium":    ("en", "en_US", "amy",      "medium", "en_US-amy-medium"),
}

DEFAULT_VOICE = os.getenv("DEFAULT_VOICE", "en_US-libritts-high")

_http = None
def http():
    global _http
    if _http is None:
        import requests
        _http = requests.Session()
        _http.headers.update({"User-Agent": "ActualTTS/CPU"})
    return _http

def _download(url: str, dest: Path, timeout: int = 300):
    dest.parent.mkdir(parents=True, exist_ok=True)
    with http().get(url, timeout=timeout, stream=True, headers={"Accept": "application/octet-stream"}) as r:
        r.raise_for_status()
        ct = (r.headers.get("content-type") or "").lower()
        if "text/html" in ct:
            raise RuntimeError(f"Bad content-type for {url}: {ct}")
        tmp = dest.with_suffix(dest.suffix + ".part")
        with open(tmp, "wb") as f:
            for chunk in r.iter_content(1 << 16):
                if chunk:
                    f.write(chunk)
        tmp.replace(dest)

def _file_ok(p: Path, min_bytes: int) -> bool:
    try:
        return p.exists() and p.stat().st_size >= min_bytes
    except Exception:
        return False

def _read_sr_from_cfg(cfg_path: Path) -> int:
    try:
        with open(cfg_path, "r", encoding="utf-8") as f:
            j = json.load(f)
        sr = int(j.get("sample_rate", 22050))
        # keep it reasonable
        return sr if sr in (16000, 22050, 24000, 44100, 48000) else 22050
    except Exception:
        return 22050

def ensure_voice(voice_id: str) -> Dict[str, object]:
    """Ensure voice .onnx and .onnx.json exist locally with sane sizes. Returns paths and SR."""
    # Aliases
    if voice_id.lower() in ("en-us", "en_us", "english"):
        voice_id = "en_US-libritts-high"
    if voice_id not in VOICE_MAP:
        raise RuntimeError(f"Unknown voice '{voice_id}'. Known: {list(VOICE_MAP)}")

    lang, country, family, quality, base = VOICE_MAP[voice_id]
    vdir  = VOICES_DIR / voice_id
    model = vdir / f"{base}.onnx"
    cfg   = vdir / f"{base}.onnx.json"

    vdir.mkdir(parents=True, exist_ok=True)

    prefix    = f"{HF_REPO_BASE}/{HF_REV}/{lang}/{country}/{family}/{quality}/{base}"
    model_url = f"{prefix}.onnx"
    cfg_url   = f"{prefix}.onnx.json"

    if not _file_ok(model, MIN_ONNX_BYTES):
        _download(model_url, model)
    if not _file_ok(cfg, MIN_JSON_BYTES):
        _download(cfg_url, cfg)

    if not _file_ok(model, MIN_ONNX_BYTES):
        sz = model.stat().st_size if model.exists() else 0
        raise RuntimeError(f"Downloaded .onnx too small ({sz} bytes) for '{voice_id}'")
    if not _file_ok(cfg, MIN_JSON_BYTES):
        sz = cfg.stat().st_size if cfg.exists() else 0
        raise RuntimeError(f"Downloaded .onnx.json too small ({sz} bytes) for '{voice_id}'")

    sr = _read_sr_from_cfg(cfg)
    return {"model": model, "config": cfg, "sr": sr}

# -------------------------
# Piper exec helpers
# -------------------------

def build_piper_cmd(
    text: str, voice_id: str, to_stdout: bool,
    out_path: Optional[Path] = None,
    length_scale: float = 1.08, noise_scale: float = 0.35, noise_w: float = 0.90
) -> list:
    vc = ensure_voice(voice_id)
    cmd = [
        *PIPER_CMD,
        "-m", str(vc["model"]),
        "-c", str(vc["config"]),
        "--length-scale", str(length_scale),
        "--noise-scale",  str(noise_scale),
        "--noise-w",      str(noise_w),
        "--sentence-silence", str(SENTENCE_SILENCE),
    ]
    if to_stdout:
        # Stream RAW PCM (16-bit little-endian)
        cmd += ["-f", "-", "--output-raw"]
    else:
        if out_path is None:
            raise ValueError("out_path required when to_stdout=False")
        # File output: Piper writes a WAV
        cmd += ["-f", str(out_path)]
    return cmd

async def piper_to_file(text, voice, out_path, length_scale, noise_scale, noise_w):
    cmd = build_piper_cmd(text, voice, to_stdout=False, out_path=out_path,
                          length_scale=length_scale, noise_scale=noise_scale, noise_w=noise_w)
    proc = await asyncio.create_subprocess_exec(
        *cmd, stdin=asyncio.subprocess.PIPE, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE
    )
    # terminate the utterance with newline
    proc.stdin.write((text + "\n").encode("utf-8"))
    await proc.stdin.drain()
    proc.stdin.close()
    await proc.wait()
    if proc.returncode != 0:
        stderr = (await proc.stderr.read()).decode("utf-8", "ignore")
        raise RuntimeError(f"Piper failed (code {proc.returncode}).\n{stderr}")

async def piper_stream_raw(
    text: str,
    voice: str,
    ws: WebSocket,
    sr: int,
    channels: int,
    length_scale: float,
    noise_scale: float,
    noise_w: float,
    prebuffer_ms: int,
    prebuffer_max_wait_ms: int,
):
    """
    Synthesize immediately; stream in *batched, clock-paced* frames:
      - Accumulate audio until `prebuffer_ms` (or `prebuffer_max_wait_ms` elapses).
      - Then send fixed batches of STREAM_BATCH_MS at a steady cadence.
    """
    cmd = build_piper_cmd(text, voice, to_stdout=True,
                          length_scale=length_scale, noise_scale=noise_scale, noise_w=noise_w)
    proc = await asyncio.create_subprocess_exec(
        *cmd, stdin=asyncio.subprocess.PIPE, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE
    )

    # Send text and close stdin
    proc.stdin.write((text + "\n").encode("utf-8"))
    await proc.stdin.drain()
    proc.stdin.close()

    async def pump_stderr():
        try:
            while True:
                line = await proc.stderr.readline()
                if not line:
                    break
                try:
                    await ws.send_text(json.dumps({"event": "log", "stderr": line.decode("utf-8", "ignore").rstrip()}))
                except Exception:
                    break
        except Exception:
            pass

    stderr_task = asyncio.create_task(pump_stderr())
    total = 0

    # framing + pacing
    bytes_per_ms = max(1, int(sr * channels * 2 / 1000))
    batch_bytes = max(bytes_per_ms, int(STREAM_BATCH_MS * bytes_per_ms))
    target_prebuffer_bytes = max(0, int(prebuffer_ms) * bytes_per_ms)

    buf = bytearray()
    started_streaming = False
    first_audio_ts = None
    pace_start_t = None
    batches_sent = 0

    try:
        while True:
            chunk = await proc.stdout.read(8192)
            if chunk:
                if first_audio_ts is None:
                    first_audio_ts = time.time()
                buf.extend(chunk)

                # Flip to streaming once prebuffer is satisfied (or the fail-safe wait elapsed)
                if not started_streaming:
                    enough = (len(buf) >= target_prebuffer_bytes) if target_prebuffer_bytes > 0 else True
                    waited = False
                    if first_audio_ts is not None and prebuffer_max_wait_ms > 0:
                        waited = ((time.time() - first_audio_ts) * 1000.0) >= prebuffer_max_wait_ms
                    if enough or waited:
                        started_streaming = True
                        pace_start_t = time.time()
                        batches_sent = 0

                # If streaming, emit batches on schedule
                if started_streaming:
                    while len(buf) >= batch_bytes:
                        due_t = pace_start_t + (batches_sent * STREAM_BATCH_MS) / 1000.0
                        sleep_s = due_t - time.time()
                        if sleep_s > 0:
                            await asyncio.sleep(sleep_s)
                        await ws.send_bytes(buf[:batch_bytes])
                        del buf[:batch_bytes]
                        total += batch_bytes
                        batches_sent += 1
                continue

            # EOF from Piper
            if not started_streaming and len(buf) > 0:
                started_streaming = True
                pace_start_t = time.time()
                batches_sent = 0

            # Flush any remaining bytes in paced batches
            while len(buf) >= batch_bytes:
                due_t = pace_start_t + (batches_sent * STREAM_BATCH_MS) / 1000.0
                sleep_s = due_t - time.time()
                if sleep_s > 0:
                    await asyncio.sleep(sleep_s)
                await ws.send_bytes(buf[:batch_bytes])
                del buf[:batch_bytes]
                total += batch_bytes
                batches_sent += 1

            # Send any tail (less than a full batch) now
            if len(buf) > 0:
                await ws.send_bytes(bytes(buf))
                total += len(buf)
                buf.clear()
            break  # done reading

        await proc.wait()
        await stderr_task

        if proc.returncode != 0:
            rem = await proc.stderr.read()
            detail = rem.decode("utf-8", "ignore").strip()
            await ws.send_text(json.dumps({"event": "error", "detail": detail or f"piper exited {proc.returncode}"}))
        else:
            if total == 0:
                await ws.send_text(json.dumps({"event": "error", "detail": "No audio produced"}))
            else:
                await ws.send_text(json.dumps({"event": "done"}))
    except WebSocketDisconnect:
        try:
            proc.kill()
        except Exception:
            pass
        try:
            await stderr_task
        except Exception:
            pass

# -------------------------
# FastAPI app & routes
# -------------------------

app = FastAPI(title="ActualTTS (CPU)")

@app.get("/health")
def health():
    voices = []
    if VOICES_DIR.exists():
        for child in VOICES_DIR.iterdir():
            if not child.is_dir():
                continue
            name = child.name
            base = child / f"{name}.onnx"
            cfg  = child / f"{name}.onnx.json"
            if _file_ok(base, MIN_ONNX_BYTES) and _file_ok(cfg, MIN_JSON_BYTES):
                voices.append({"id": name, "sr": _read_sr_from_cfg(cfg)})
    # optional environment versions
    try:
        import numpy, onnxruntime as ort
        numpy_version = numpy.__version__
        onnxruntime_version = ort.__version__
    except Exception:
        numpy_version = onnxruntime_version = None

    return {
        "ok": True,
        "engine": "piper-tts (CLI, CPU)",
        "default_voice": DEFAULT_VOICE,
        "voice_dir": str(VOICES_DIR),
        "available_voices": voices,
        "files_dir": str(FILES_DIR),
        "numpy": numpy_version,
        "onnxruntime": onnxruntime_version,
    }

@app.get("/")
def root():
    return PlainTextResponse("ActualTTS (CPU) — use POST /speak, GET/POST /speak.wav, or WS /ws/tts")

@app.post("/provision")
async def provision(request: Request, x_auth: Optional[str] = Header(None)):
    """
    POST JSON: { "voice": "en_US-amy-medium" }
    Downloads voice assets if missing. Returns {ok, voice, sr}.
    """
    if not _auth_ok(x_auth):
        return JSONResponse({"ok": False, "error": "unauthorized"}, status_code=401)
    try:
        body = await request.json()
    except Exception:
        return JSONResponse({"ok": False, "error": "invalid json"}, status_code=400)
    voice = (body.get("voice") or DEFAULT_VOICE).strip()
    try:
        info = ensure_voice(voice)
        return {"ok": True, "voice": voice, "sr": int(info.get("sr", 22050))}
    except Exception as e:
        return JSONResponse({"ok": False, "error": str(e)}, status_code=500)

@app.get("/file/{name}")
def get_file(name: str):
    path = FILES_DIR / name
    if not path.exists():
        return JSONResponse({"ok": False, "error": "not found"}, status_code=404)
    return FileResponse(path)

def _validate_text(text: str) -> Optional[str]:
    if not text:
        return "Missing text"
    if len(text) > MAX_TEXT_CHARS:
        return f"text too long (>{MAX_TEXT_CHARS} chars)"
    return None

@app.post("/speak")
async def speak(request: Request, x_auth: Optional[str] = Header(None)):
    """
    POST JSON:
      { "text": "Hello", "voice": "en_US-libritts-high",
        "length_scale": 1.08, "noise_scale": 0.35, "noise_w": 0.90 }
    Returns: { "ok": true, "audio_url": "/file/tts-XXXX.wav" }
    """
    if not _auth_ok(x_auth):
        return JSONResponse({"ok": False, "error": "unauthorized"}, status_code=401)
    try:
        body = await request.json()
    except Exception:
        return JSONResponse({"detail": "Invalid JSON"}, status_code=400)

    text = (body.get("text") or "").strip()
    err = _validate_text(text)
    if err:
        return JSONResponse({"detail": err}, status_code=400)

    voice        = (body.get("voice") or DEFAULT_VOICE).strip()
    length_scale = float(body.get("length_scale", 1.08))
    noise_scale  = float(body.get("noise_scale", 0.35))
    noise_w      = float(body.get("noise_w", 0.90))

    ts = int(time.time() * 1000)
    out_path = FILES_DIR / f"tts-{ts}.wav"

    try:
        ensure_voice(voice)
        await piper_to_file(text, voice, out_path, length_scale, noise_scale, noise_w)
    except Exception as e:
        return JSONResponse({"ok": False, "error": str(e)}, status_code=500)

    return {"ok": True, "audio_url": f"/file/{out_path.name}"}

@app.post("/speak.wav")
async def speak_wav_post(request: Request, background_tasks: BackgroundTasks, x_auth: Optional[str] = Header(None)):
    """POST JSON -> returns audio/wav directly"""
    if not _auth_ok(x_auth):
        return JSONResponse({"ok": False, "error": "unauthorized"}, status_code=401)
    try:
        body = await request.json()
    except Exception:
        return JSONResponse({"detail": "Invalid JSON"}, status_code=400)

    text = (body.get("text") or "").strip()
    err = _validate_text(text)
    if err:
        return JSONResponse({"detail": err}, status_code=400)

    voice        = (body.get("voice") or DEFAULT_VOICE).strip()
    length_scale = float(body.get("length_scale", 1.08))
    noise_scale  = float(body.get("noise_scale", 0.35))
    noise_w      = float(body.get("noise_w", 0.90))

    ts = int(time.time() * 1000)
    out_path = FILES_DIR / f"tts-{ts}.wav"

    try:
        ensure_voice(voice)
        await piper_to_file(text, voice, out_path, length_scale, noise_scale, noise_w)
    except Exception as e:
        return JSONResponse({"ok": False, "error": str(e)}, status_code=500)

    background_tasks.add_task(_safe_unlink, out_path)
    return FileResponse(out_path, media_type="audio/wav", filename=out_path.name, background=background_tasks)

@app.get("/speak.wav")
async def speak_wav_get(
    text: str,
    voice: str = DEFAULT_VOICE,
    length_scale: float = 1.08,
    noise_scale: float = 0.35,
    noise_w: float = 0.90,
    background_tasks: BackgroundTasks = None,
    x_auth: Optional[str] = Header(None),
):
    """GET query -> returns audio/wav directly"""
    if not _auth_ok(x_auth):
        return JSONResponse({"ok": False, "error": "unauthorized"}, status_code=401)

    text = (text or "").strip()
    err = _validate_text(text)
    if err:
        return JSONResponse({"detail": err}, status_code=400)

    ts = int(time.time() * 1000)
    out_path = FILES_DIR / f"tts-{ts}.wav"

    try:
        ensure_voice(voice.strip())
        await piper_to_file(text, voice.strip(), out_path, float(length_scale), float(noise_scale), float(noise_w))
    except Exception as e:
        return JSONResponse({"ok": False, "error": str(e)}, status_code=500)

    if background_tasks:
        background_tasks.add_task(_safe_unlink, out_path)
    return FileResponse(out_path, media_type="audio/wav", filename=out_path.name)

# --- Diagnostics: inspect/refresh downloaded voices ---
@app.get("/debug/voices")
def debug_voices(redownload: bool = Query(False, description="Force re-download bad/missing files")):
    out = {"dir": str(VOICES_DIR), "voices": []}
    for vid, (_lang, _country, _family, _quality, base) in VOICE_MAP.items():
        vdir = VOICES_DIR / vid
        model = vdir / f"{base}.onnx"
        cfg   = vdir / f"{base}.onnx.json"
        info = {
            "id": vid,
            "model": str(model), "cfg": str(cfg),
            "model_exists": model.exists(), "cfg_exists": cfg.exists(),
            "model_size": (model.stat().st_size if model.exists() else 0),
            "cfg_size": (cfg.stat().st_size if cfg.exists() else 0),
            "sr": _read_sr_from_cfg(cfg) if cfg.exists() else None,
        }
        out["voices"].append(info)

        if redownload and (not _file_ok(model, MIN_ONNX_BYTES) or not _file_ok(cfg, MIN_JSON_BYTES)):
            try:
                vdir.mkdir(parents=True, exist_ok=True)
                for p in vdir.glob("*"):
                    p.unlink(missing_ok=True)
                ensure_voice(vid)
                info["redownloaded"] = True
                info["model_size"] = (model.stat().st_size if model.exists() else 0)
                info["cfg_size"] = (cfg.stat().st_size if cfg.exists() else 0)
                info["sr"] = _read_sr_from_cfg(cfg) if cfg.exists() else None
            except Exception as e:
                info["redownload_error"] = str(e)
    return out

# --- Live streaming WS ---
@app.websocket("/ws/tts")
async def ws_tts(ws: WebSocket):
    await ws.accept()
    voice = DEFAULT_VOICE
    length_scale, noise_scale, noise_w = 1.08, 0.35, 0.90
    voice_sr = 22050  # will be set from config on init
    # NEW: defaults (client can override in init)
    prebuffer_ms = PREBUFFER_MS
    prebuffer_max_wait_ms = PREBUFFER_MAX_WAIT_MS

    try:
        while True:
            msg = await ws.receive_text()
            try:
                data = json.loads(msg)
            except Exception:
                continue
            ev = data.get("event")
            if ev == "init":
                # optional shared-secret over WS: accept via token field
                token = (data.get("token") or "")
                if AUTH_SHARED_SECRET and token != AUTH_SHARED_SECRET:
                    await ws.send_text(json.dumps({"event": "error", "detail": "unauthorized"}))
                    await ws.close(); return

                voice = (data.get("voice") or voice).strip()
                # Accept explicit params first
                if "length_scale" in data: length_scale = float(data["length_scale"])
                if "noise_scale"  in data: noise_scale  = float(data["noise_scale"])
                if "noise_w"      in data: noise_w      = float(data["noise_w"])
                # Optional: map rate_wpm → length_scale if user didn't set a custom length_scale
                if "length_scale" not in data and "rate_wpm" in data:
                    try:
                        rate_wpm = int(data.get("rate_wpm", 165))
                        length_scale = max(0.70, min(1.40, 165.0 / max(100, rate_wpm)))
                    except Exception:
                        pass
                # NEW: allow client override for prebuffer knobs
                if "prebuffer_ms" in data:
                    try: prebuffer_ms = max(0, int(data["prebuffer_ms"]))
                    except Exception: pass
                if "prebuffer_max_wait_ms" in data:
                    try: prebuffer_max_wait_ms = max(0, int(data["prebuffer_max_wait_ms"]))
                    except Exception: pass

                try:
                    info = ensure_voice(voice)
                    voice_sr = int(info.get("sr", 22050))
                except Exception as e:
                    await ws.send_text(json.dumps({"event": "error", "detail": str(e)}))
                    await ws.close()
                    return
                # Announce the correct SR so the client opens the device properly
                await ws.send_text(json.dumps({"event": "ready", "sr": voice_sr, "channels": DEFAULT_CH}))
            elif ev == "speak":
                text = (data.get("text") or "").strip()
                if not text:
                    await ws.send_text(json.dumps({"event": "error", "detail": "empty text"}))
                    continue
                if len(text) > MAX_TEXT_CHARS:
                    await ws.send_text(json.dumps({"event":"error","detail": f"text too long (>{MAX_TEXT_CHARS})"}))
                    continue
                await piper_stream_raw(
                    text, voice, ws, voice_sr, DEFAULT_CH, length_scale, noise_scale, noise_w,
                    prebuffer_ms, prebuffer_max_wait_ms
                )
            # ignore others
    except WebSocketDisconnect:
        return
    except Exception as e:
        try:
            await ws.send_text(json.dumps({"event": "error", "detail": str(e)}))
        except Exception:
            pass
        try:
            await ws.close()
        except Exception:
            pass

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=int(os.getenv("PORT", "7860")), reload=False)