Fabian
84 lines
3.3 KiB
Python
84 lines
3.3 KiB
Python
"""Qwen3-TTS HTTP server — OpenAI-compatible /v1/audio/speech endpoint."""
|
|
import argparse, io, os, sys
|
|
from flask import Flask, request, send_file, jsonify
|
|
|
|
app = Flask(__name__)
|
|
MODEL_DIR = "/models/qwen3-tts"
|
|
engine = None
|
|
|
|
SPEAKERS = ["Vivian","Serena","uncle_fu","Dylan","Eric","Ryan","Aiden","ono_anna","Sohee"]
|
|
LANGUAGES = ["english","chinese","japanese","korean","german","spanish","french","russian","italian","portuguese"]
|
|
|
|
def init_engine():
|
|
global engine
|
|
# TTSEngine uses Path.cwd() as project_root and builds relative paths from it.
|
|
# model_dir must be a subdirectory of cwd for relative_to() to work.
|
|
parent = os.path.dirname(os.path.abspath(MODEL_DIR))
|
|
basename = os.path.basename(MODEL_DIR)
|
|
os.chdir(parent)
|
|
from qwen3_tts_gguf.inference import TTSEngine
|
|
engine = TTSEngine(model_dir=basename, onnx_provider="CPUExecutionProvider")
|
|
if not engine.ready:
|
|
print(f"[server] WARNING: engine created but not ready", flush=True)
|
|
|
|
@app.route("/health")
|
|
def health():
|
|
ready = engine is not None and engine.ready
|
|
return jsonify({"status": "ok" if ready else "loading", "speakers": SPEAKERS, "languages": LANGUAGES})
|
|
|
|
@app.route("/v1/audio/speech", methods=["POST"])
|
|
def speech():
|
|
if engine is None or not engine.ready:
|
|
return jsonify({"error": "engine not ready"}), 503
|
|
data = request.get_json(force=True)
|
|
text = data.get("input", "")
|
|
if not text:
|
|
return jsonify({"error": "missing input"}), 400
|
|
speaker = data.get("voice", "Vivian")
|
|
if speaker not in SPEAKERS:
|
|
speaker = "Vivian"
|
|
language = data.get("language", "english")
|
|
if language not in LANGUAGES:
|
|
language = "english"
|
|
instruct = data.get("instruct", "")
|
|
try:
|
|
import numpy as np, wave as _wave
|
|
from qwen3_tts_gguf.inference import TTSConfig
|
|
stream = engine.create_stream(n_ctx=2048)
|
|
if stream is None:
|
|
return jsonify({"error": "failed to create stream"}), 500
|
|
cfg = TTSConfig()
|
|
result = stream.custom(text=text, speaker=speaker, language=language,
|
|
instruct=instruct, config=cfg)
|
|
stream.join()
|
|
if result is None:
|
|
return jsonify({"error": "synthesis returned None"}), 500
|
|
audio = result.audio if hasattr(result, 'audio') and result.audio is not None else None
|
|
if audio is None:
|
|
tmp = os.path.join("/tmp", "tts_out.wav")
|
|
result.save(tmp)
|
|
import soundfile as sf
|
|
audio, _ = sf.read(tmp, dtype='float32')
|
|
os.remove(tmp)
|
|
buf = io.BytesIO()
|
|
with _wave.open(buf, "wb") as wf:
|
|
wf.setnchannels(1)
|
|
wf.setsampwidth(2)
|
|
wf.setframerate(24000)
|
|
pcm = (np.clip(audio, -1.0, 1.0) * 32767).astype(np.int16).tobytes()
|
|
wf.writeframes(pcm)
|
|
buf.seek(0)
|
|
return send_file(buf, mimetype="audio/wav", download_name="speech.wav")
|
|
except Exception as e:
|
|
return jsonify({"error": str(e)}), 500
|
|
|
|
if __name__ == "__main__":
|
|
p = argparse.ArgumentParser()
|
|
p.add_argument("--host", default="0.0.0.0")
|
|
p.add_argument("--port", type=int, default=8072)
|
|
p.add_argument("--model-dir", default=MODEL_DIR)
|
|
a = p.parse_args()
|
|
MODEL_DIR = a.model_dir
|
|
init_engine()
|
|
app.run(host=a.host, port=a.port)
|