Add reproducible Piper TTS service

2026-08-21 17:02:32 +02:00
parent 53dfffc289
commit a5cf11582a
19 changed files with 316 additions and 34 deletions
@@ -0,0 +1,153 @@
+#!/usr/bin/env python3
+"""Small, private Piper worker for the Mike AI profile router.
+
+The public OpenAI-compatible endpoint remains in the router.  This worker only
+accepts the narrow internal /status and /tts protocol and never logs input text.
+"""
+
+from __future__ import annotations
+
+import io
+import json
+import os
+import subprocess
+import threading
+import wave
+from http import HTTPStatus
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+from pathlib import Path
+
+from piper import PiperVoice, SynthesisConfig
+
+
+DATA_DIR = Path(os.getenv("PIPER_DATA_DIR", "/data"))
+VOICE_NAME = os.getenv("PIPER_VOICE", "de_DE-thorsten-high")
+VOICE_ALIAS = os.getenv("PIPER_VOICE_ALIAS", "alloy")
+HOST = os.getenv("PIPER_HOST", "0.0.0.0")
+PORT = int(os.getenv("PIPER_PORT", "8085"))
+MAX_TEXT_CHARS = int(os.getenv("PIPER_MAX_TEXT_CHARS", "8000"))
+MAX_REQUEST_BYTES = int(os.getenv("PIPER_MAX_REQUEST_BYTES", "65536"))
+
+VOICE_PATH = DATA_DIR / f"{VOICE_NAME}.onnx"
+VOICE = PiperVoice.load(str(VOICE_PATH))
+SYNTHESIS_LOCK = threading.Lock()
+
+
+def synthesize_wav(text: str, speed: float) -> bytes:
+    """Synthesize a complete WAV in memory without retaining the text."""
+    output = io.BytesIO()
+    config = SynthesisConfig(length_scale=1.0 / speed)
+    with SYNTHESIS_LOCK, wave.open(output, "wb") as wav_file:
+        VOICE.synthesize_wav(text, wav_file, syn_config=config)
+    return output.getvalue()
+
+
+def wav_to_mp3(wav_bytes: bytes) -> bytes:
+    """Convert Piper's WAV to the MP3 format Open WebUI requests by default."""
+    result = subprocess.run(
+        [
+            "ffmpeg", "-hide_banner", "-loglevel", "error",
+            "-f", "wav", "-i", "pipe:0",
+            "-codec:a", "libmp3lame", "-b:a", "96k",
+            "-f", "mp3", "pipe:1",
+        ],
+        input=wav_bytes,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+        check=False,
+        timeout=120,
+    )
+    if result.returncode != 0:
+        raise RuntimeError("ffmpeg conversion failed")
+    return result.stdout
+
+
+class Handler(BaseHTTPRequestHandler):
+    protocol_version = "HTTP/1.1"
+
+    def log_message(self, fmt: str, *args: object) -> None:
+        # Deliberately omit URLs and request bodies from the log.
+        print(f"piper-worker: {self.command} -> {args[1] if len(args) > 1 else '-'}")
+
+    def send_bytes(self, status: int, body: bytes, content_type: str) -> None:
+        self.send_response(status)
+        self.send_header("Content-Type", content_type)
+        self.send_header("Content-Length", str(len(body)))
+        self.send_header("Cache-Control", "no-store")
+        self.end_headers()
+        self.wfile.write(body)
+
+    def send_json(self, status: int, payload: dict) -> None:
+        self.send_bytes(
+            status,
+            json.dumps(payload, separators=(",", ":")).encode(),
+            "application/json",
+        )
+
+    def do_GET(self) -> None:  # noqa: N802
+        if self.path != "/status":
+            self.send_json(HTTPStatus.NOT_FOUND, {"error": "not found"})
+            return
+        self.send_json(
+            HTTPStatus.OK,
+            {
+                "ready": True,
+                "engine": "piper",
+                "model": VOICE_NAME,
+                "voices": [VOICE_ALIAS],
+            },
+        )
+
+    def do_POST(self) -> None:  # noqa: N802
+        if self.path != "/tts":
+            self.send_json(HTTPStatus.NOT_FOUND, {"error": "not found"})
+            return
+
+        try:
+            content_length = int(self.headers.get("Content-Length", "0"))
+        except ValueError:
+            content_length = 0
+        if content_length <= 0 or content_length > MAX_REQUEST_BYTES:
+            self.send_json(HTTPStatus.REQUEST_ENTITY_TOO_LARGE, {"error": "invalid request size"})
+            return
+
+        try:
+            request = json.loads(self.rfile.read(content_length))
+            text = request.get("text", "")
+            voice = request.get("voice", VOICE_ALIAS)
+            output_format = request.get("format", "mp3")
+            speed = float(request.get("speed", 1.0))
+        except (json.JSONDecodeError, TypeError, ValueError):
+            self.send_json(HTTPStatus.BAD_REQUEST, {"error": "invalid JSON request"})
+            return
+
+        if not isinstance(text, str) or not text.strip() or len(text) > MAX_TEXT_CHARS:
+            self.send_json(HTTPStatus.BAD_REQUEST, {"error": "invalid text"})
+            return
+        if voice != VOICE_ALIAS:
+            self.send_json(HTTPStatus.BAD_REQUEST, {"error": "unknown voice"})
+            return
+        if output_format not in {"wav", "mp3"}:
+            self.send_json(HTTPStatus.BAD_REQUEST, {"error": "unsupported format"})
+            return
+        if not 0.5 <= speed <= 2.0:
+            self.send_json(HTTPStatus.BAD_REQUEST, {"error": "invalid speed"})
+            return
+
+        try:
+            audio = synthesize_wav(text.strip(), speed)
+            if output_format == "mp3":
+                audio = wav_to_mp3(audio)
+                content_type = "audio/mpeg"
+            else:
+                content_type = "audio/wav"
+        except (OSError, RuntimeError, subprocess.SubprocessError):
+            self.send_json(HTTPStatus.INTERNAL_SERVER_ERROR, {"error": "synthesis failed"})
+            return
+
+        self.send_bytes(HTTPStatus.OK, audio, content_type)
+
+
+if __name__ == "__main__":
+    print(f"Piper worker ready: {VOICE_NAME} as {VOICE_ALIAS} on {HOST}:{PORT}")
+    ThreadingHTTPServer((HOST, PORT), Handler).serve_forever()