python: speech to text offline

识别音频声音 ，转换不是很准

# encoding: utf-8
# 版权所有  2026 ©涂聚文有限公司™ ®
# 许可信息查看：言語成了邀功盡責的功臣，還需要行爲每日來值班嗎
# 描述：pip install moviepy SpeechRecognition  pip install openai-whisper
# Author    : geovindu,Geovin Du 涂聚文.
# IDE       : PyCharm 2024.3.6 python 3.11
# os        : windows 10
# database  : mysql 9.0 sql server 2019, postgreSQL 17.0  Oracle 21c Neo4j
# Datetime  : 2026/6/14 17:40
# User      :  geovindu
# Product   : PyCharm
# Project   : Pysimple
# File      : MP4totextdest.py
 
import warnings
warnings.filterwarnings("ignore")
warnings.filterwarnings("ignore", category=RuntimeWarning)
 
 
from moviepy import VideoFileClip
import speech_recognition as sr
import os
import whisper
import numpy as np
import wave
import audioop
import whisper.audio
 
 
VIDEO_FILE = "20260614_102306.mp4"
TEMP_WAV = "temp_audio.wav"
OUTPUT_TXT = "完整演讲文稿_带时间戳.txt"
 
# 1. 提取视频音频
print("正在提取音频...")
video = VideoFileClip(VIDEO_FILE)
# 导出16位单声道标准wav，适配识别
video.audio.write_audiofile(TEMP_WAV, codec="pcm_s16le", fps=16000)
video.close()
 
 
# 2. 原生Python读取wav，彻底绕过ffmpeg
def load_audio_without_ffmpeg(path):
    with wave.open(path, 'rb') as wav_file:
        n_channels = wav_file.getnchannels()
        width = wav_file.getsampwidth()
        frames = wav_file.readframes(wav_file.getnframes())
 
    # 双声道转单声道
    if n_channels == 2:
        frames = audioop.tomono(frames, width, 0.5, 0.5)
    # 转浮点音频数组
    audio_np = np.frombuffer(frames, dtype=np.int16).flatten().astype(np.float32) / 32768.0
    return audio_np
 
 
# 覆盖whisper自带读取函数，不再调用ffmpeg
import whisper.audio
 
whisper.audio.load_audio = load_audio_without_ffmpeg
 
# 3. 加载最轻量模型 tiny（速度最快，中文演讲够用）
print("加载离线语音模型 tiny ...")
model = whisper.load_model("tiny")
 
# 4. 开始识别，开启分段时间戳
print("开始逐段识别演讲内容，请等待...")
result = model.transcribe(
    audio=TEMP_WAV,
    language="zh",
    verbose=False,
    word_timestamps=True
)
 
# 5. 格式化带时间戳文稿，实时打印每一段
full_content = "=== EV录屏 第十八届海峡论坛苏恒演讲 完整转写稿 ===\n\n"
print("\n====================识别结果====================")
for seg in result["segments"]:
    start_min = int(seg["start"] // 60)
    start_sec = int(seg["start"] % 60)
    end_min = int(seg["end"] // 60)
    end_sec = int(seg["end"] % 60)
    seg_text = seg["text"].strip()
 
    line = f"[{start_min:02d}:{start_sec:02d} - {end_min:02d}:{end_sec:02d}] {seg_text}"
    print(line)  # 实时控制台打印每一段文字
    full_content += line + "\n"
 
# 6. 保存本地文本
with open(OUTPUT_TXT, "w", encoding="utf-8") as f:
    f.write(full_content)
 
# 清理临时音频文件
os.remove(TEMP_WAV)
print(f"\n================================================")
print(f"✅ 全部识别完成！文稿已保存至：{OUTPUT_TXT}")
　　

输出：

# 超高准确率版（适合演讲、新闻、正式发言）
model = whisper.load_model("medium")  

# encoding: utf-8 
# 版权所有  2026 ©涂聚文有限公司™ ®
# 许可信息查看：言語成了邀功盡責的功臣，還需要行爲每日來值班嗎
# 描述：pip install moviepy SpeechRecognition  pip install openai-whisper
# Author    : geovindu,Geovin Du 涂聚文.
# IDE       : PyCharm 2024.3.6 python 3.11
# os        : windows 10
# database  : mysql 9.0 sql server 2019, postgreSQL 17.0  Oracle 21c Neo4j
# Datetime  : 2026/6/14 17:40 
# User      :  geovindu
# Product   : PyCharm
# Project   : Pysimple
# File      : MP4totextdest.py

import warnings
warnings.filterwarnings("ignore")
warnings.filterwarnings("ignore", category=RuntimeWarning)


from moviepy import VideoFileClip
import speech_recognition as sr
import os
import whisper
import numpy as np
import wave
import audioop
import whisper.audio


VIDEO_FILE = "geovindu.mp4"
TEMP_WAV = "temp_audio.wav"
OUTPUT_TXT = "完整演讲文稿_带时间戳medium.txt"

# 1. 提取视频音频
print("正在提取音频...")
video = VideoFileClip(VIDEO_FILE)
# 导出16位单声道标准wav，适配识别
video.audio.write_audiofile(TEMP_WAV, codec="pcm_s16le", fps=16000)
video.close()


# 2. 原生Python读取wav，彻底绕过ffmpeg
def load_audio_without_ffmpeg(path):
    with wave.open(path, 'rb') as wav_file:
        n_channels = wav_file.getnchannels()
        width = wav_file.getsampwidth()
        frames = wav_file.readframes(wav_file.getnframes())

    # 双声道转单声道
    if n_channels == 2:
        frames = audioop.tomono(frames, width, 0.5, 0.5)
    # 转浮点音频数组
    audio_np = np.frombuffer(frames, dtype=np.int16).flatten().astype(np.float32) / 32768.0
    return audio_np


# 覆盖whisper自带读取函数，不再调用ffmpeg
import whisper.audio

whisper.audio.load_audio = load_audio_without_ffmpeg

# 3. 加载最轻量模型 tiny（速度最快，中文演讲够用）
print("加载离线语音模型 medium ...")
model = whisper.load_model("medium")

# 4. 开始识别，开启分段时间戳
print("开始逐段识别演讲内容，请等待...")
result = model.transcribe(
    audio=TEMP_WAV,
    language="zh",
    verbose=False,
    word_timestamps=True
)

# 5. 格式化带时间戳文稿，实时打印每一段
full_content = "=== EV录屏 2026年6月13日 第十八届海峡论坛苏恒演讲 完整转写稿 ===\n\n"
print("\n====================识别结果====================")
for seg in result["segments"]:
    start_min = int(seg["start"] // 60)
    start_sec = int(seg["start"] % 60)
    end_min = int(seg["end"] // 60)
    end_sec = int(seg["end"] % 60)
    seg_text = seg["text"].strip()

    line = f"[{start_min:02d}:{start_sec:02d} - {end_min:02d}:{end_sec:02d}] {seg_text}"
    print(line)  # 实时控制台打印每一段文字
    full_content += line + "\n"

# 6. 保存本地文本
with open(OUTPUT_TXT, "w", encoding="utf-8") as f:
    f.write(full_content)

# 清理临时音频文件
os.remove(TEMP_WAV)
print(f"\n================================================")
print(f"✅ 全部识别完成！文稿已保存至：{OUTPUT_TXT}")

输出：