notes-qpid-pro/app/tts.py

import os
import re
import hashlib
import asyncio
import logging
from pathlib import Path
from typing import Optional, Dict, Any

from app import config

logger = logging.getLogger("notes_app.tts")

PIPER_DIR = config.BASE_DIR / "data" / "piper"
PIPER_BIN = PIPER_DIR / "piper"
MODELS_DIR = PIPER_DIR / "models"
AUDIO_CACHE_DIR = config.BASE_DIR / "data" / "audio_cache"

# Ensure audio cache directory exists
AUDIO_CACHE_DIR.mkdir(parents=True, exist_ok=True)

VOICE_MODELS = {
    "en": {
        "model": MODELS_DIR / "en_US-lessac-medium.onnx",
        "config": MODELS_DIR / "en_US-lessac-medium.onnx.json",
        "name": "en_US-lessac",
    },
    "ru": {
        "model": MODELS_DIR / "ru_RU-dmitri-medium.onnx",
        "config": MODELS_DIR / "ru_RU-dmitri-medium.onnx.json",
        "name": "ru_RU-dmitri",
    },
}


def clean_tts_text(text: str) -> str:
    """Cleans input text from markdown, HTML tags, and abnormal whitespace."""
    if not text:
        return ""
    # Remove HTML tags if any
    t = re.sub(r"<[^>]+>", " ", text)
    # Remove markdown link formatting [text](url) -> text
    t = re.sub(r"\[([^\]]+)\]\([^)]+\)", r"\1", t)
    # Collapse multiple whitespaces
    t = " ".join(t.strip().split())
    return t


def get_audio_hash(text: str, lang: str = "en", speed: float = 1.0) -> str:
    """Computes a deterministic MD5 hash for audio caching."""
    clean = clean_tts_text(text).lower()
    norm_speed = round(float(speed), 2)
    norm_lang = "ru" if lang.lower().startswith("ru") else "en"
    key = f"{norm_lang}:{norm_speed:.2f}:{clean}".encode("utf-8")
    return hashlib.md5(key).hexdigest()


def get_cached_audio_path(text: str, lang: str = "en", speed: float = 1.0) -> Optional[Path]:
    """Returns Path if already cached and valid, else None."""
    h = get_audio_hash(text, lang, speed)
    file_path = AUDIO_CACHE_DIR / f"{h}.wav"
    if file_path.is_file() and file_path.stat().st_size > 100:
        return file_path
    return None


def is_piper_available() -> bool:
    """Checks if piper binary and at least en model exist and are executable."""
    if not PIPER_BIN.is_file() or not os.access(PIPER_BIN, os.X_OK):
        return False
    en_model = VOICE_MODELS["en"]["model"]
    return en_model.is_file()


async def synthesize_speech(
    text: str,
    lang: str = "en",
    speed: float = 1.0,
) -> Optional[Path]:
    """
    Synthesizes speech using local Piper TTS and returns Path to the generated WAV file.
    Fast disk-cached: if already generated, returns immediately without running synthesis.
    """
    clean = clean_tts_text(text)
    if not clean:
        return None

    # Check cache first
    cached_path = get_cached_audio_path(clean, lang, speed)
    if cached_path:
        return cached_path

    if not is_piper_available():
        logger.warning("Piper TTS binary or models not available")
        return None

    norm_lang = "ru" if lang.lower().startswith("ru") else "en"
    voice_info = VOICE_MODELS.get(norm_lang, VOICE_MODELS["en"])
    model_path = voice_info["model"]
    config_path = voice_info["config"]

    if not model_path.is_file() or not config_path.is_file():
        logger.warning(f"Voice model files missing for {norm_lang}: {model_path}")
        return None

    h = get_audio_hash(clean, norm_lang, speed)
    out_wav = AUDIO_CACHE_DIR / f"{h}.wav"
    temp_wav = AUDIO_CACHE_DIR / f"{h}.tmp.wav"

    # Piper length_scale: > 1.0 is slower, < 1.0 is faster
    # 0.75x speed -> length_scale ~ 1.33
    # 1.0x speed  -> length_scale = 1.0
    # 1.25x speed -> length_scale ~ 0.8
    clamped_speed = max(0.5, min(2.0, float(speed)))
    length_scale = str(round(1.0 / clamped_speed, 2))

    env = os.environ.copy()
    ld_path = str(PIPER_DIR)
    if "LD_LIBRARY_PATH" in env and env["LD_LIBRARY_PATH"]:
        env["LD_LIBRARY_PATH"] = f"{ld_path}:{env['LD_LIBRARY_PATH']}"
    else:
        env["LD_LIBRARY_PATH"] = ld_path

    cmd = [
        str(PIPER_BIN),
        "--model", str(model_path),
        "--config", str(config_path),
        "--espeak_data", str(PIPER_DIR / "espeak-ng-data"),
        "--length_scale", length_scale,
        "--output_file", str(temp_wav),
    ]

    try:
        proc = await asyncio.create_subprocess_exec(
            *cmd,
            stdin=asyncio.subprocess.PIPE,
            stdout=asyncio.subprocess.PIPE,
            stderr=asyncio.subprocess.PIPE,
            env=env,
        )
        stdout, stderr = await proc.communicate(input=clean.encode("utf-8"))

        if proc.returncode == 0 and temp_wav.is_file() and temp_wav.stat().st_size > 100:
            temp_wav.replace(out_wav)
            logger.info(f"Synthesized speech [{norm_lang}, speed={speed}]: {clean[:40]}... -> {out_wav.name}")
            return out_wav
        else:
            err_text = stderr.decode("utf-8", errors="replace").strip()
            logger.warning(f"Piper exited code {proc.returncode}: {err_text}")
            if temp_wav.exists():
                temp_wav.unlink(missing_ok=True)
    except Exception as e:
        logger.warning(f"Piper execution error: {e}")
        if temp_wav.exists():
            temp_wav.unlink(missing_ok=True)

    return None


async def prewarm_card_audio(term: str, context: Optional[str] = None):
    """
    Background pre-synthesis of word and context sentence.
    Invoked when a user saves a card to ensure zero-latency audio playback in SRS.
    """
    try:
        if term and term.strip():
            await synthesize_speech(term.strip(), lang="en", speed=1.0)
        if context and context.strip():
            clean_ctx = context.strip().replace('"', '').strip()
            # If context is very long, cap to first sentence or 300 chars
            if len(clean_ctx) > 300:
                clean_ctx = clean_ctx[:300]
            await synthesize_speech(clean_ctx, lang="en", speed=1.0)
    except Exception as e:
        logger.warning(f"Background prewarm error: {e}")