enginex-mlu370-tts/mlu_370-kokoro/kokoro_server.py

import os
import io

from fastapi import FastAPI, Response, Body, HTTPException
from fastapi.responses import StreamingResponse, JSONResponse
from contextlib import asynccontextmanager
import uvicorn
import xml.etree.ElementTree as ET

from kokoro import KPipeline, KModel
# import soundfile as sf
import wave
import numpy as np
from scipy.signal import resample

import torch
from torch import Tensor
from torch.nn import functional as F
from typing import Optional, List
import re
from dataclasses import dataclass

import kokoro.istftnet as _ist

def _inverse_no_complex(self, magnitude, phase):
    """
    解决 MLU 设备上不支持复数计算
    """
    device = magnitude.device
    dtype = magnitude.dtype

    win_dev = torch.hann_window(self.win_length, device=device, dtype=dtype)

    real = magnitude * torch.cos(phase)
    imag = magnitude * torch.sin(phase)
    spec_ri = torch.stack([real, imag], dim=-1).contiguous()  # (..., 2)

    real_cpu = real.to("cpu")
    imag_cpu = imag.to("cpu")
    spec_complex_cpu = torch.complex(real_cpu, imag_cpu)  # (..,) 复数张量
    win_cpu = torch.hann_window(self.win_length, device="cpu", dtype=dtype)

    wav_cpu = torch.istft(
        spec_complex_cpu,
        n_fft=self.filter_length,
        hop_length=self.hop_length,
        win_length=self.win_length,
        window=win_cpu,
        center=True,
        normalized=False,
        onesided=True,
    )
    return wav_cpu.to(device).unsqueeze(-2)

def _transform_no_complex(self, input_data):
    """
    纯实数 STFT：return_complex=False，随后手动求幅度与相位
    """
    z = torch.stft(
        input_data,
        n_fft=self.filter_length,
        hop_length=self.hop_length,
        win_length=self.win_length,
        window=self.window.to(input_data.device, dtype=input_data.dtype),
        return_complex=False,
        center=True,
        normalized=False,
    )
    real = z[..., 0]
    imag = z[..., 1]
    magnitude = torch.sqrt(real * real + imag * imag)
    phase = torch.atan2(imag, real)
    return magnitude, phase

# 替换 Kokoro 的 STFT.inverse 实现
_ist.TorchSTFT.inverse = _inverse_no_complex
_ist.TorchSTFT.transform = _transform_no_complex


repo_id = 'hexgrad/Kokoro-82M-v1.1-zh'
MODEL_SR = 24000
TARGET_SR = 16000
# How much silence to insert between paragraphs: 5000 is about 0.2 seconds
N_ZEROS = 20
model = None
en_empty_pipeline = None
en_voice = os.getenv('EN_VOICE', 'af_maple.pt')
zh_voice = os.getenv('ZH_VOICE', 'zf_046.pt')
model_dir = os.getenv('MODEL_DIR', '/model/hexgrad')
model_name = os.getenv('MODEL_NAME','kokoro-v1_1-zh.pth')

# model_1_1_dir = os.path.join(model_dir, 'Kokoro-82M-v1.1-zh')
# model_1_0_dir = os.path.join(model_dir, 'Kokoro-82M')
# repo_id_1_0 = 'hexgrad/Kokoro-82M'

@dataclass
class LanguagePipeline:
    pipeline: KPipeline
    voice_pt: str

pipeline_dict: dict[str, LanguagePipeline] = {}

def en_callable(text):
    if text == 'Kokoro':
        return 'kˈOkəɹO'
    elif text == 'Sol':
        return 'sˈOl'
    return next(en_empty_pipeline(text)).phonemes

# HACK: Mitigate rushing caused by lack of training data beyond ~100 tokens
# Simple piecewise linear fn that decreases speed as len_ps increases
def speed_callable(len_ps):
    speed = 0.8
    if len_ps <= 83:
        speed = 1
    elif len_ps < 183:
        speed = 1 - (len_ps - 83) / 500
    return speed


# from https://huggingface.co/spaces/coqui/voice-chat-with-mistral/blob/main/app.py
def wave_header_chunk(frame_input=b"", channels=1, sample_width=2, sample_rate=32000):
    # This will create a wave header then append the frame input
    # It should be first on a streaming wav file
    # Other frames better should not have it (else you will hear some artifacts each chunk start)
    wav_buf = io.BytesIO()
    with wave.open(wav_buf, "wb") as vfout:
        vfout.setnchannels(channels)
        vfout.setsampwidth(sample_width)
        vfout.setframerate(sample_rate)
        vfout.writeframes(frame_input)

    wav_buf.seek(0)
    return wav_buf.read()


def resample_audio(data: np.ndarray, original_rate: int, target_rate: int):
    ori_dtype = data.dtype
    # data = normalize_audio(data)
    number_of_samples = int(len(data) * float(target_rate) / original_rate)
    resampled_data = resample(data, number_of_samples)
    # resampled_data = normalize_audio(resampled_data)
    return resampled_data.astype(ori_dtype)

def audio_postprocess(data: np.ndarray, original_rate: int, target_rate: int):
    audio = resample_audio(data, original_rate, target_rate)
    if audio.dtype == np.float32:
        audio = np.int16(audio * 32767)
        audio = np.concatenate([audio, np.zeros(N_ZEROS, dtype=np.int16)])
    return audio

def init():
    global model, en_empty_pipeline
    global model_1_0
    global pipeline_dict
    # device = 'cuda' if torch.cuda.is_available() else 'cpu'
    device = 'mlu'
    model = KModel(repo_id=repo_id, model=os.path.join(model_dir, model_name), config=os.path.join(model_dir, 'config.json')).to(device).eval()
    en_empty_pipeline = KPipeline(lang_code='a', repo_id=repo_id, model=False)
    en_pipeline = KPipeline(lang_code='a', repo_id=repo_id, model=model)
    zh_pipeline = KPipeline(lang_code='z', repo_id=repo_id, model=model, en_callable=en_callable)
    en_voice_pt = os.path.join(model_dir, 'voices', en_voice)
    zh_voice_pt = os.path.join(model_dir, 'voices', zh_voice)
    pipeline_dict['zh'] = LanguagePipeline(pipeline=zh_pipeline, voice_pt=zh_voice_pt)
    pipeline_dict['en'] = LanguagePipeline(pipeline=en_pipeline, voice_pt=en_voice_pt)

    # v1.0 model for other languages
    # model_1_0 = KModel(repo_id=repo_id_1_0, model=os.path.join(model_1_0_dir, 'kokoro-v1_0.pth'), config=os.path.join(model_1_0_dir, 'config.json')).to(device).eval()
    # # es
    # es_pipeline = KPipeline(lang_code='e', repo_id=repo_id_1_0, model=model_1_0)
    # es_voice_pt = os.path.join(model_1_0_dir, 'voices', 'ef_dora.pt')
    # pipeline_dict['es'] = LanguagePipeline(pipeline=es_pipeline, voice_pt=es_voice_pt)
    # # fr
    # fr_pipeline = KPipeline(lang_code='f', repo_id=repo_id_1_0, model=model_1_0)
    # fr_voice_pt = os.path.join(model_1_0_dir, 'voices', 'ff_siwis.pt')
    # pipeline_dict['fr'] = LanguagePipeline(pipeline=fr_pipeline, voice_pt=fr_voice_pt)
    # # hi
    # hi_pipeline = KPipeline(lang_code='h', repo_id=repo_id_1_0, model=model_1_0)
    # hi_voice_pt = os.path.join(model_1_0_dir, 'voices', 'hf_alpha.pt')
    # pipeline_dict['hi'] = LanguagePipeline(pipeline=hi_pipeline, voice_pt=hi_voice_pt)
    # # it
    # it_pipeline = KPipeline(lang_code='i', repo_id=repo_id_1_0, model=model_1_0)
    # it_voice_pt = os.path.join(model_1_0_dir, 'voices', 'if_sara.pt')
    # pipeline_dict['it'] = LanguagePipeline(pipeline=it_pipeline, voice_pt=it_voice_pt)
    # # ja
    # ja_pipeline = KPipeline(lang_code='j', repo_id=repo_id_1_0, model=model_1_0)
    # ja_voice_pt = os.path.join(model_1_0_dir, 'voices', 'jf_alpha.pt')
    # pipeline_dict['ja'] = LanguagePipeline(pipeline=ja_pipeline, voice_pt=ja_voice_pt)
    # # pt
    # pt_pipeline = KPipeline(lang_code='p', repo_id=repo_id_1_0, model=model_1_0)
    # pt_voice_pt = os.path.join(model_1_0_dir, 'voices', 'pf_dora.pt')
    # pipeline_dict['pt'] = LanguagePipeline(pipeline=pt_pipeline, voice_pt=pt_voice_pt)

    warmup()

@asynccontextmanager
async def lifespan(app: FastAPI):
    init()
    yield
    pass

app = FastAPI(lifespan=lifespan)

def warmup():
    zh_pipeline = pipeline_dict['zh'].pipeline
    voice = pipeline_dict['zh'].voice_pt
    generator = zh_pipeline(text="语音合成测试TTS。", voice=voice, speed=speed_callable)
    for _ in generator:
        pass

xml_namespace = "{http://www.w3.org/XML/1998/namespace}"

symbols = ',.!?;:()[]{}<>，。！？；：【】《》……"“”_—'
def contains_words(text):
    return any(char not in symbols for char in text)

def cut_sentences(text) -> list[str]:
    text = text.strip()
    splits = re.split(r"([.;?!、。？！；])", text)
    sentences = []
    for i in range(0, len(splits), 2):
        if i + 1 < len(splits):
            s = splits[i] + splits[i + 1]
        else:
            s = splits[i]
        s = s.strip()
        if s:
            sentences.append(s)
    return sentences

LANGUAGE_ALIASES = {
    'z': 'zh',
    'a': 'en',
    'e': 'es',
    'f': 'fr',
    'h': 'hi',
    'i': 'it',
    'j': 'ja',
    'p': 'pt',
}

@app.post("/")
@app.post("/tts")
def predict(ssml: str = Body(...), include_header: bool = False):
    try:
        root = ET.fromstring(ssml)
        voice_element = root.find(".//voice")
        if voice_element is not None:
            transcription = voice_element.text.strip()
            language = voice_element.get(f'{xml_namespace}lang', "zh").strip()
            # voice_name = voice_element.get("name", "zh-f-soft-1").strip()
        else:
            return JSONResponse(status_code=400, content={"message": "Invalid SSML format: <voice> element not found."})
    except ET.ParseError as e:
        return JSONResponse(status_code=400, content={"message": "Invalid SSML format", "Exception": str(e)})
    
    if not contains_words(transcription):
        audio = np.zeros(N_ZEROS, dtype=np.int16).tobytes()
        if include_header:
            audio_header = wave_header_chunk(sample_rate=TARGET_SR)
            audio = audio_header + audio
        return Response(audio, media_type='audio/wav')
    
    if language not in pipeline_dict:
        if language in LANGUAGE_ALIASES:
            language = LANGUAGE_ALIASES[language]
        else:
            return JSONResponse(status_code=400, content={"message": f"Language '{language}' not supported."})


    def streaming_generator():
        texts = cut_sentences(transcription)
        has_yield = False
        for text in texts:
            if text.strip() and contains_words(text):
                pipeline = pipeline_dict[language].pipeline
                voice = pipeline_dict[language].voice_pt
                if language == 'zh':
                    generator = pipeline(text=text, voice=voice, speed=speed_callable)
                else:
                    generator = pipeline(text=text, voice=voice)

                for (_, _, audio) in generator:
                    if include_header and not has_yield:
                        has_yield = True
                        yield wave_header_chunk(sample_rate=TARGET_SR)
                    yield audio_postprocess(audio.numpy(), MODEL_SR, TARGET_SR).tobytes()

    return StreamingResponse(streaming_generator(), media_type='audio/wav')


@app.get("/health")
@app.get("/ready")
async def ready():
    return JSONResponse(status_code=200, content={"status": "ok"})

@app.get("/health_check")
async def health_check():
    try:
        a = torch.ones(10, 20, dtype=torch.float32, device='cuda')
        b = torch.ones(20, 10, dtype=torch.float32, device='cuda')
        c = torch.matmul(a, b)
        if c.sum() == 10 * 20 * 10:
            return {"status": "ok"}
        else:
            raise HTTPException(status_code=503)
    except Exception as e:
        print(f'health_check failed')
        raise HTTPException(status_code=503)

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=80)