Whisper+T5-translate实现python实时语音翻译

最新推荐文章于 2025-07-13 18:11:32 发布

原创最新推荐文章于 2025-07-13 18:11:32 发布 · 623 阅读

3 ·

CC 4.0 BY-SA版权

文章标签：

#whisper #python #开发语言

1.首先下载模型，加载模型

import torch
import numpy as np
import webrtcvad
import pyaudio
import queue
import threading
from datetime import datetime
from faster_whisper import WhisperModel
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM,pipeline
from transformers import T5ForConditionalGeneration, T5Tokenizer
device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16
save_directory = "./faster-distil-whiper-large-v3-local"  # 替换为你希望保存的本地路径
# en_zh_directory = "./opus-mt-en-zh-local"  # 替换为你希望保存的本地路径
en_zh_directory = "./t5-translate-en-ru-zh-base-200-sent-local"  # 替换为你希望保存的本地路径
whisperModel = WhisperModel(save_directory, device="cuda", compute_type="float32")


model = T5ForConditionalGeneration.from_pretrained(en_zh_directory)
model.eval()
model.to(device)
tokenizer = T5Tokenizer.from_pretrained(en_zh_directory)
vad = webrtcvad.Vad(3)  # 设置 VAD 灵敏度（0-3，3 最敏感）
prefix = 'translate to zh: '

2.配置麦克风

# 初始化 PyAudio
p = pyaudio.PyAudio()
# 设置音频流参数
FORMAT = pyaudio.paInt16  # 16-bit 音频格式
CHANNELS = 1              # 单声道
RATE = 16000              # 采样率（Whisper 需要 16kHz）
FRAME_DURATION = 20       # 每帧的时长（ms）
CHUNK = int(RATE * FRAME_DURATION / 1000)  # 每帧的帧数
MIN_SILENCE_DURATION = 0.2  # 最小静音时长（秒）

3.队列构建，构建录音基本参数

# 共享队列，用于录音和推理线程之间的数据交换
audio_queue = queue.Queue()

silence_frames = 0
silence_frames_lock = threading.Lock()

4.构建录音函数

# 录音线程
def record_audio():
    global silence_frames
    stream = p.open(
        format=FORMAT,
        channels=CHANNELS,
        rate=RATE,
        input=True,
        frames_per_buffer=CHUNK,
    )
    print("开始录音...按 Ctrl+C 停止")
    try:
        while True:
            # 从麦克风读取音频数据
            data = stream.read(CHUNK)
            audio_data = np.frombuffer(data, dtype=np.int16).astype(np.float32) / 32768.0

            # 使用 VAD 检测语音活动
            if vad.is_speech(data, RATE):
                audio_queue.put(audio_data)
                with silence_frames_lock:
                    silence_frames = 0  # 重置静音计数器
            else:
                with silence_frames_lock:
                    silence_frames += 1  # 重置静音计数器
    except KeyboardInterrupt:
        print("录音停止")
    finally:
        stream.stop_stream()
        stream.close()
        p.terminate()

5.构建翻译函数

def process_audio():
    global silence_frames
    audio_buffer = np.array([], dtype=np.float32)
    silence_frames = 0

    while True:
        try:
            # 从队列中获取音频数据
            audio_data = audio_queue.get(timeout=1)  # 超时 1 秒
            audio_buffer = np.concatenate((audio_buffer, audio_data))

        except Exception as e:
            pass
        current_silence_frames = 0
        # 检查静音计数器
        with silence_frames_lock:
            current_silence_frames = silence_frames

        # 如果检测到静音时间超过阈值，处理累积的音频
        if (current_silence_frames > MIN_SILENCE_DURATION * (RATE / CHUNK)) or len(audio_buffer) > 320 * 200:
            if(len(audio_buffer) > 0):#, language="en"
                segments, _ = whisperModel.transcribe(audio_buffer,vad_filter=True,vad_parameters=dict(min_silence_duration_ms=200), language="en", condition_on_previous_text=True)
                for segment in segments:
                    if(segment.text == ""):
                        continue
                    elif(segment.text == "Thank you."):
                        print("[%s] %s (%s)" % (str(datetime.now()), "感谢", segment.text))
                    else:            
                        src_text = prefix + segment.text
                        
                        input_ids = tokenizer(src_text, return_tensors="pt")
                        generated_tokens = model.generate(**input_ids.to(device))
                        result = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
                        print("[%s] %s (%s)" % (str(datetime.now()), result[0], segment.text))
                    
                    
                    # result = pipeline(segment.text)
                    # print("[%s] %s (%s)" % (str(datetime.now()), result[0]['translation_text'], segment.text))
            
                audio_buffer = np.array([], dtype=np.float16)
                silence_frames = 0

6.启动线程，启动程序

# 启动录音线程和推理线程
record_thread = threading.Thread(target=record_audio)
process_thread = threading.Thread(target=process_audio)

record_thread.start()
process_thread.start()