Alfredv2/alfred/app/services/piper_tts.py

import asyncio
import os
import subprocess
import tempfile
import time
import threading  # ALFRED_DUPLEX_BARGE_IN_V1
import wave

from wyoming.audio import AudioChunk, AudioStart, AudioStop
from wyoming.client import AsyncTcpClient
from wyoming.tts import Synthesize, SynthesizeVoice


PIPER_HOST = os.getenv("PIPER_HOST", "127.0.0.1")
PIPER_PORT = int(os.getenv("PIPER_PORT", "10200"))
AUDIO_DEVICE = os.getenv("AUDIO_OUTPUT_DEVICE", "plughw:1,0")


async def synthesize(text, wav_path):
    client = AsyncTcpClient(PIPER_HOST, PIPER_PORT)

    await client.connect()

    try:
        await client.write_event(
            Synthesize(
                       text=text,
                       voice=SynthesizeVoice(
                       name=os.getenv("PIPER_VOICE", "fr_FR-tom-medium")
                       )).event()
        )

        wav_file = None

        try:
            while True:
                event = await asyncio.wait_for(
                    client.read_event(),
                    timeout=30
                )

                if event is None:
                    raise RuntimeError("Connexion Piper interrompue")

                if AudioStart.is_type(event.type):
                    start = AudioStart.from_event(event)

                    wav_file = wave.open(wav_path, "wb")
                    wav_file.setnchannels(start.channels)
                    wav_file.setsampwidth(start.width)
                    wav_file.setframerate(start.rate)

                elif AudioChunk.is_type(event.type):
                    chunk = AudioChunk.from_event(event)

                    if wav_file is None:
                        wav_file = wave.open(wav_path, "wb")
                        wav_file.setnchannels(chunk.channels)
                        wav_file.setsampwidth(chunk.width)
                        wav_file.setframerate(chunk.rate)

                    wav_file.writeframes(chunk.audio)

                elif AudioStop.is_type(event.type):
                    break

        finally:
            if wav_file is not None:
                wav_file.close()

    finally:
        await client.disconnect()


# ALFRED_DUPLEX_BARGE_IN_V1: control only this module's aplay process.
_playback_lock = threading.RLock()
_playback_cancel = threading.Event()
_playback_proc = None


def stop_playback():
    """Interrupts Piper playback or marks pending synth for cancellation."""
    global _playback_proc
    _playback_cancel.set()
    with _playback_lock:
        process = _playback_proc
    if process is not None and process.poll() is None:
        try:
            process.terminate()
        except OSError:
            pass


# DONA_WAVES_V1 — volume réel du WAV Wyoming Piper (50 ms, hors micro).
from array import array as _dona_array
import sys as _dona_sys
_dona_level_callback = None

def dona_set_level_callback(callback):
    global _dona_level_callback
    _dona_level_callback = callback

def _dona_emit_level(value):
    if _dona_level_callback is not None:
        try:
            _dona_level_callback(max(0.0, min(1.0, float(value))))
        except Exception:
            pass  # Une erreur côté interface ne doit jamais interrompre Piper.

def _dona_waves_stream_envelope(wav_path, proc):
    """Échantillonne le PCM pendant aplay, sans enregistrer ni exposer le son."""
    try:
        with wave.open(wav_path, 'rb') as stream:
            if stream.getsampwidth() != 2:
                return  # Autre format : état « speaking » en secours côté navigateur.
            rate=stream.getframerate()
            if rate <= 0:
                return
            frames=max(1,int(rate * .05))
            started=time.monotonic()
            index=0
            while proc.poll() is None:
                raw=stream.readframes(frames)
                if not raw:
                    break
                samples=_dona_array('h')
                samples.frombytes(raw)
                if _dona_sys.byteorder=='big':
                    samples.byteswap()
                stride=max(1,len(samples)//420)
                window=samples[::stride]
                rms=(sum(s*s for s in window)/max(1,len(window)))**.5
                desired=started+index*frames/rate
                remaining=desired-time.monotonic()
                if remaining>0:
                    time.sleep(min(remaining,.07))
                if proc.poll() is not None:
                    break
                _dona_emit_level(min(1.0,rms/6500))
                index+=1
    except Exception as exc:
        print('[DONA WAVES] analyse PCM impossible :',type(exc).__name__,flush=True)
    finally:
        _dona_emit_level(0)

def speak(text):
    if not text or not text.strip():
        return

    _playback_cancel.clear()  # ALFRED_DUPLEX_BARGE_IN_V1

    fd, wav_path = tempfile.mkstemp(
        prefix="alfred_piper_",
        suffix=".wav"
    )
    os.close(fd)

    start = time.monotonic()

    try:
        print("[ALFRED PIPER] Synthèse...", flush=True)

        asyncio.run(synthesize(text, wav_path))

        print("[ALFRED PIPER] Lecture...", flush=True)

        # ALFRED_DUPLEX_BARGE_IN_V1: allow a concurrent microphone detector
        # to terminate playback, without killing Hermes or the entire service.
        global _playback_proc
        if _playback_cancel.is_set():
            print("[ALFRED PIPER] Annulé avant lecture", flush=True)
            return
        proc = subprocess.Popen(["aplay", "-q", "-D", AUDIO_DEVICE, wav_path])
        with _playback_lock:
            _playback_proc = proc
            if _playback_cancel.is_set() and proc.poll() is None:
                proc.terminate()
        # DONA_WAVES_V1 : animation par l'amplitude du vrai son joué.
        threading.Thread(target=_dona_waves_stream_envelope,
                         args=(wav_path,proc),daemon=True).start()
        try:
            result = proc.wait(timeout=120)
        except subprocess.TimeoutExpired:
            proc.kill()
            proc.wait()
            raise
        finally:
            with _playback_lock:
                if _playback_proc is proc:
                    _playback_proc = None
        if _playback_cancel.is_set():
            print("[ALFRED PIPER] Lecture interrompue par la voix", flush=True)
            return
        if result:
            raise subprocess.CalledProcessError(result, proc.args)

        print(
            f"[ALFRED PIPER] Terminé en {time.monotonic() - start:.2f}s",
            flush=True
        )

    except Exception as e:
        print(f"[ALFRED PIPER] Erreur : {e}", flush=True)
        raise
    finally:
        if os.path.exists(wav_path):
            os.remove(wav_path)