PPASR: 基于PaddlePaddle的中文语音识别框架¶

PPASR
├── docs                  # 文档
├── download_data         # 数据下载与处理
├── dataset               # 数据集相关
├── models                # 模型保存路径
├── lm                    # 语言模型
├── logs                  # 日志
├── utils                 # 工具函数
├── train.py              # 训练脚本
├── eval.py               # 评估脚本
├── export_model.py       # 模型导出
├── infer_path.py         # 路径预测
├── infer_server.py       # 服务预测
├── infer_gui.py          # GUI预测
└── README.md             # 项目说明

pip install -r requirements.txt

from ppasr.predict import PPASRPredictor

# 初始化预测器，使用预训练模型
predictor = PPASRPredictor(model_tag='conformer_streaming_fbank_wenetspeech')

# 短语音识别
wav_path = 'test.wav'
result = predictor.predict(audio_data=wav_path, use_pun=False)
score, text = result['score'], result['text']
print(f"识别结果: {text}, 得分: {int(score)}")

# 长语音识别
result = predictor.predict_long(audio_data=wav_path, use_pun=False)
score, text = result['score'], result['text']
print(f"长语音识别结果: {text}, 得分: {score}")

# 实时流式识别
import time
import wave

wf = wave.open(wav_path, 'rb')
CHUNK = int(16000 * 0.5)  # 每0.5秒一个chunk
data = wf.readframes(CHUNK)
while data != b'':
    start = time.time()
    result = predictor.predict_stream(audio_data=data, use_pun=False, is_end=data == b'')
    data = wf.readframes(CHUNK)
    if result:
        score, text = result['score'], result['text']
        print(f"实时结果: {text}, 得分: {int(score)}")
        print(f"耗时: {int((time.time() - start) * 1000)}ms")
predictor.reset_stream()

python export_model.py --resume_model=models/conformer_online_fbank/best_model/

cd download_data
python download.py  # 下载Aishell、Free ST-Chinese-Mandarin-Corpus、THCHS-30

dataset/audio/001.wav  这是一段测试语音
dataset/audio/002.wav  语音识别测试

python create_data.py

python train.py

python -m paddle.distributed.launch --gpus '0,1' train.py

python eval.py --resume_model=models/conformer_online_fbank/best_model/

python export_model.py --resume_model=models/conformer_online_fbank/best_model/

python infer_path.py --wav_path=test.wav

python infer_server.py

import requests
import wave

wav_data = wave.open('test.wav', 'rb').readframes(-1)
response = requests.post(
    "http://localhost:8080/transcribe",
    files={"audio": ("test.wav", wav_data)}
)
print(response.json())

python infer_gui.py