这个版本基本完成了VAD+SAR

这个提交包含在:
2025-09-07 23:09:12 +08:00
父节点 0c9c883700
当前提交 5da91c7303
共有 2 个文件被更改,包括 91 次插入 和 54 次删除

128
R2T.py
查看文件

@ -4,61 +4,81 @@ import sherpa_onnx
from config import Config from config import Config
# 配置参数 sample_rate = 16000 # 必须与VAD和识别器配置的采样率匹配
sample_rate = 16000 block_size = 512 # 每次回调处理的样本数,较小的值延迟低,但CPU占用高
is_recording = False
audio_data = []
def audio_callback(indata, frames, time_info, status): vadconfig = sherpa_onnx.VadModelConfig()
"""音频回调函数""" vadconfig.silero_vad.model = Config.vad
# global audio_data vadconfig.silero_vad.threshold = 0.5
audio_data.extend(indata[:, 0].copy()) vadconfig.silero_vad.min_silence_duration = 0.5 # seconds
vadconfig.silero_vad.min_speech_duration = 0.25 # seconds
vadconfig.silero_vad.max_speech_duration = 8 # seconds
vadconfig.sample_rate = sample_rate
vad = sherpa_onnx.VoiceActivityDetector(vadconfig,buffer_size_in_seconds=85)
is_recording = False
samples = np.array([], dtype=np.float32) # 用于累积一个语音段的样本
recognizer = sherpa_onnx.OfflineRecognizer.from_paraformer(
tokens = Config.tokens,
paraformer= Config.paraformer
)
def audio_callback(indata, frames, time, status):
global is_recording, samples
# indata 是一个 (frames, channels) 的 numpy 数组
# 我们假设是单声道,并将其转换为一维数组
audio_data = indata[:, 0].astype(np.float32)
# 1. 将音频数据送入 VAD
vad.accept_waveform(audio_data)
# 2. 检查当前音频块是否有语音
if vad.is_speech_detected():
if not is_recording:
# 语音刚开始,重置累积的样本
print("Speech started, recording...")
samples = np.array([], dtype=np.float32)
is_recording = True
# 无论是否刚开始,有语音就累积数据
samples = np.concatenate((samples, audio_data))
else:
# 当前块没有检测到语音
if is_recording:
# 之前正在录制,说明语音段刚刚结束
print("Speech ended. Decoding...")
is_recording = False
# 3. 将累积的整个语音段样本传入识别器进行解码
# 注意:sherpa_onnx的OfflineRecognizer使用decode()处理整个数组
stream = recognizer.create_stream()
stream.accept_waveform(sample_rate=16000, waveform=samples)
recognizer.decode_stream(stream)
result = stream.result.text
print(f"Recognized: {result}")
# 可选:清空累积的样本以备下次使用
samples = np.array([], dtype=np.float32)
def Record2Text(): def Record2Text():
global is_recording, audio_data pass
# 创建识别器
recognizer = sherpa_onnx.OfflineRecognizer.from_paraformer(
tokens = Config.tokens,
paraformer= Config.paraformer
)
try:
print("1")
# 打开音频流
stream = sd.InputStream(
samplerate=sample_rate,
channels=1,
dtype='float32',
blocksize=1024,
callback=audio_callback
)
with stream:
audio_data = []
print("录音中... 请说话")
input("按回车结束录音...")
print("处理中...")
# 处理录音数据
if audio_data:
s = recognizer.create_stream()
s.accept_waveform(sample_rate, np.array(audio_data, dtype=np.float32))
recognizer.decode_stream(s)
result = s.result
if result and result.text and result.text.strip():
print(f"\n识别结果: {result.text}")
return result.text
else:
print("\n未识别到有效语音")
return ""
else:
print("\n没有录制到音频")
return ""
except KeyboardInterrupt: try:
print("\n程序退出") with sd.InputStream(callback=audio_callback,
channels=1,
except Exception as e: samplerate=sample_rate,
print(f"发生错误: {e}") blocksize=block_size,
dtype='float32'):
print("Recording started. Press Enter to stop.")
input() # 阻塞,等待用户输入以停止
except KeyboardInterrupt:
print("用户输入中断")
print("Recording stopped.")
if __name__ == '__main__':
Record2Text()

17
R2Test.py 普通文件
查看文件

@ -0,0 +1,17 @@
import sounddevice as sd
import numpy as np
import sherpa_onnx
from config import Config
recognizer = sherpa_onnx.OfflineRecognizer.from_paraformer(
tokens = Config.tokens,
paraformer= Config.paraformer
)
InS = sd.InputStream(
samplerate=16000,
channels=1,
dtype='float32',
blocksize=1024
)