init ascend tts

2025-09-05 11:27:43 +08:00
parent d53ac91bb6
commit b92a65b0fa
602 changed files with 590901 additions and 1 deletions
--- a/ascend_910-matcha/Dockerfile_matcha
+++ b/ascend_910-matcha/Dockerfile_matcha
@@ -0,0 +1,19 @@
+FROM quay.io/ascend/vllm-ascend:v0.10.0rc1
+
+WORKDIR /workspace
+
+RUN cp /etc/apt/sources.list /etc/apt/sources.list.bak && \
+    echo "deb [trusted=yes] https://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ jammy main restricted universe multiverse" > /etc/apt/sources.list && \
+    echo "deb [trusted=yes] https://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ jammy-updates main restricted universe multiverse" >> /etc/apt/sources.list && \
+    echo "deb [trusted=yes] https://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ jammy-backports main restricted universe multiverse" >> /etc/apt/sources.list && \
+    echo "deb [trusted=yes] https://mirrors.tuna.tsinghua.edu.cn/ubuntu-ports/ jammy-security main restricted universe multiverse" >> /etc/apt/sources.list && \
+    apt-get update && \
+    apt-get install -y espeak-ng && \
+    rm -rf /var/lib/apt/lists/*
+
+COPY requirements_matcha.txt constraints_matcha.txt matcha_server.py launch_matcha.sh /workspace/
+
+RUN pip install -r requirements_matcha.txt -c constraints_matcha.txt 
+RUN pip install matcha-tts -c constraints_matcha.txt
+
+ENTRYPOINT ["/bin/bash", "launch_matcha.sh"]
--- a/ascend_910-matcha/README.md
+++ b/ascend_910-matcha/README.md
@@ -0,0 +1,53 @@
+# Matcha-TTS
+
+本项目基于 **matcha** 模型封装，提供简洁的 Docker 部署方式，支持 **SSML 输入**，输出 **PCM 原始音频**，可用于语音合成。
+
+---
+
+## Quickstart
+
+### 1. 安装镜像
+```bash
+docker build -t tts:matcha . -f Dockerfile_matcha
+```
+
+### 2. 启动服务
+```bash
+docker run -it --rm \
+  -v /models/matcha_model:/mnt/models \
+  -e ASCEND_VISIBLE_DEVICES=1 \
+  --device /dev/davinci2:/dev/davinci0 \
+  --device /dev/davinci_manager \
+  --device /dev/devmm_svm \
+  --device /dev/hisi_hdc \
+  -v /usr/local/dcmi:/usr/local/dcmi \
+  -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
+  -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
+  -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
+  -v /etc/ascend_install.info:/etc/ascend_install.info \
+  --privileged \
+  -p 8080:80 \
+  -e MODEL_DIR=/mnt/models \
+  -e MODEL_NAME=model.ckpt \
+  tts:matcha
+```
+
+参数说明：
+- `MODEL_DIR`：模型所在目录（挂载到容器内 `/mnt/models`）
+- `MODEL_NAME`：加载的模型文件名（通常为 `.safetensors`）
+- `-p 8080:80`：将容器内服务端口映射到宿主机 `8080`
+
+### 3. 测试服务
+```bash
+curl --request POST "http://localhost:8080/tts" \
+  --header 'Content-Type: application/ssml+xml' \
+  --header 'User-Agent: curl' \
+  --data-raw '<speak version="1.0" xml:lang="zh">
+    <voice xml:lang="zh" xml:gender="Female" name="zh">
+      今天天气很好，不知道明天天气怎么样。
+    </voice>
+  </speak>' \
+  --output sound.pcm
+```
+
+---
--- a/ascend_910-matcha/constraints_matcha.txt
+++ b/ascend_910-matcha/constraints_matcha.txt
@@ -0,0 +1 @@
+torch==2.7.1
--- a/ascend_910-matcha/launch_matcha.sh
+++ b/ascend_910-matcha/launch_matcha.sh
@@ -0,0 +1,4 @@
+#!/bin/bash
+
+python3 matcha_server.py
+
--- a/ascend_910-matcha/matcha_server.py
+++ b/ascend_910-matcha/matcha_server.py
@@ -0,0 +1,182 @@
+import os
+model_dir = os.getenv("MODEL_DIR", "/mounted_model")
+model_name = os.getenv("MODEL_NAME", "model.ckpt")
+
+import logging
+logging.basicConfig(
+    format="%(asctime)s %(name)-12s %(levelname)-4s %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+    level=os.environ.get("LOGLEVEL", "INFO"),
+)
+logger = logging.getLogger(__file__)
+
+import wave
+import numpy as np
+from scipy.signal import resample
+import re
+
+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
+import torch 
+
+torch.set_default_dtype(torch.float32)
+
+
+_original_hann_window = torch.hann_window
+
+def _safe_hann_window(window_length,
+                      periodic=True,
+                      *,
+                      dtype=None,
+                      layout=torch.strided,
+                      device=None,
+                      requires_grad=False,
+                      **kwargs):
+    """
+    NPU 不支持int64 hann_window, 替换实现
+    """
+    if dtype is None:
+        dtype = torch.float32
+
+    # 总是在 CPU 先生成，绕过 NPU 上的 in-place cos 实现
+    win = _original_hann_window(
+        window_length,
+        periodic=periodic,
+        dtype=dtype,
+        layout=layout,
+        device="cpu",
+        requires_grad=requires_grad,
+        **kwargs,
+    )
+
+    if device is not None:
+        win = win.to(device)
+
+    return win
+
+torch.hann_window = _safe_hann_window
+
+from torch import Tensor
+from torch.nn import functional as F
+from typing import Optional, List
+
+from matcha.cli import load_matcha, load_vocoder, to_waveform, process_text
+
+model = None
+vocoder = None
+denoiser = None
+device = 'npu'
+
+MODEL_SR = int(os.getenv("MODEL_SR", 22050))
+speaking_rate = float(os.getenv("SPEAKING_RATE", 1.0))
+TARGET_SR = 16000
+N_ZEROS = 100
+
+
+def init():
+    global model, vocoder, denoiser
+    ckpt_path = os.path.join(model_dir, model_name)
+    vocoder_path = os.path.join(model_dir, "generator_v1")
+    model = load_matcha("custom_model", ckpt_path, device)
+    vocoder, denoiser = load_vocoder("hifigan_T2_v1", vocoder_path, device)
+
+    # warmup:
+    for _ in generate("你好，欢迎使用语音合成服务。"):
+        pass
+
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    init()
+    yield
+    pass
+
+app = FastAPI(lifespan=lifespan)
+
+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 split_text(text, max_chars=135):
+    sentences = re.split(r"(?<=[;:.!?])\s+|(?<=[。！？])", text)
+    sentences = [s.strip() for s in sentences if s.strip()]
+    return sentences
+
+def audio_postprocess(audio: np.ndarray, ori_sr: int, target_sr: int) -> np.ndarray:
+    if ori_sr != target_sr:
+        number_of_samples = int(len(audio) * float(target_sr) / ori_sr)
+        audio_resampled = resample(audio, number_of_samples)
+    else:
+        audio_resampled = audio
+    if audio.dtype == np.float32:
+        audio_resampled = np.clip(audio_resampled, -1.0, 1.0)
+        audio_resampled = (audio_resampled * 32767).astype(np.int16)
+    return audio_resampled
+
+def generate(texts):
+    chunks = split_text(texts)
+    for i, chunk in enumerate(chunks):
+        try:
+            text_processed = process_text(0, chunk, device)
+        except Exception as e:
+            logger.error(f"Error processing text: {e}")
+        with torch.inference_mode():
+            output = model.synthesise(
+                text_processed["x"],
+                text_processed["x_lengths"],
+                n_timesteps=10,
+                temperature=0.667,
+                spks=None,
+                length_scale=speaking_rate
+            )
+            output["waveform"] = to_waveform(output["mel"], vocoder, denoiser, denoiser_strength=0.00025)
+            audio = output["waveform"].detach().cpu().squeeze().numpy()
+        yield audio_postprocess(audio, MODEL_SR, TARGET_SR).tobytes()
+
+
+@app.post("/")
+@app.post("/tts")
+def predict(ssml: str = Body(...)):
+    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()
+        return Response(audio, media_type='audio/wav')
+    
+    return StreamingResponse(generate(transcription), media_type='audio/wav')
+
+@app.get("/health")
+@app.get("/ready")
+async def ready():
+    return JSONResponse(status_code=200, content={"message": "success"})
+
+@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__":
+    import uvicorn
+    uvicorn.run(app, host="0.0.0.0", port=80)
--- a/ascend_910-matcha/requirements_matcha.txt
+++ b/ascend_910-matcha/requirements_matcha.txt
@@ -0,0 +1,47 @@
+# --------- pytorch --------- #
+torch>=2.0.0
+torchvision>=0.15.0
+lightning>=2.0.0
+torchmetrics>=0.11.4
+
+# --------- hydra --------- #
+hydra-core==1.3.2
+hydra-colorlog==1.2.0
+hydra-optuna-sweeper==1.2.0
+
+# --------- loggers --------- #
+# wandb
+# neptune-client
+# mlflow
+# comet-ml
+# aim>=3.16.2  # no lower than 3.16.2, see https://github.com/aimhubio/aim/issues/2550
+
+# --------- others --------- #
+rootutils       # standardizing the project root setup
+pre-commit      # hooks for applying linters on commit
+rich            # beautiful text formatting in terminal
+pytest          # tests
+# sh            # for running bash commands in some tests (linux/macos only)
+phonemizer      # phonemization of text
+tensorboard
+librosa
+Cython
+numpy
+einops
+inflect
+Unidecode
+scipy
+torchaudio
+matplotlib
+pandas
+conformer==0.3.2
+diffusers # developed using version ==0.25.0
+notebook
+ipywidgets
+gradio==3.43.2
+gdown
+wget
+seaborn
+
+fastapi
+uvicorn[standard]