diora-web/tts/piper_engine.py

import io
import threading
import wave

from django.conf import settings

DEFAULT_LANGUAGE = 'de'
SUPPORTED_LANGUAGES = tuple(settings.TTS_VOICES.keys())

# Lazy per-process singletons, one per language: each gunicorn worker loads a
# voice only once it's actually requested, rather than all workers loading
# every model at startup (the host runs several other containers with limited
# spare RAM). One lock guards both the lazy-load and the inference call below
# — a single self-hosted user never needs concurrent synthesis across
# languages, so there's no reason for a lock per voice.
_voices = {}
_lock = threading.Lock()


def _get_voice(lang):
    if lang not in _voices:
        with _lock:
            if lang not in _voices:
                from piper import PiperVoice
                _voices[lang] = PiperVoice.load(str(settings.TTS_VOICES[lang]))
    return _voices[lang]


def synthesize_wav(text, lang=DEFAULT_LANGUAGE):
    """Synthesize `text` (in `lang`) to WAV bytes.

    Never persists or logs `text` — callers must not log it either. The lock
    also serializes inference, since one onnxruntime session isn't meant to
    run concurrent calls within a process.
    """
    voice = _get_voice(lang)
    buf = io.BytesIO()
    with _lock:
        with wave.open(buf, 'wb') as wav_file:
            voice.synthesize_wav(text, wav_file)
    return buf.getvalue()