ASR — Speech-to-Text (FunASR + Whisper)#

Two engines for different scenarios:

EngineBest ForChinese QualitySpeedFunASR SenseVoice (default)Chinese, Japanese, Korean⭐⭐⭐ 简体Fast (0.03 RTF)OpenAI WhisperMultilingual, translation⭐⭐ (繁体)Slower

Quick Start#

bash

# Default: FunASR SenseVoice (best Chinese)
{baseDir}/scripts/asr.py --input audio.mp3

# Whisper for multilingual / translation
{baseDir}/scripts/asr.py --input audio.mp3 --engine whisper

Options#

OptionDefaultDescription--input(required)Input audio file (mp3, wav, m4a, etc.)--enginefunasrASR engine: funasr (SenseVoice) or whisper``--languageautoLanguage code: zh, en, ja, ko, etc. (auto-detect if omitted)--modelbaseWhisper model size: tiny/base/small/medium/large (whisper only)--tasktranscribetranscribe or translate (whisper only)--outputWrite transcript to file (default: stdout)

Engine Details#

FunASR SenseVoice-Small (Default)#

• Model: iic/SenseVoiceSmall (893MB, auto-downloaded from ModelScope)
• Strengths: 简体中文最佳、情感识别、语音事件检测、速度极快
• Output: 简体中文，自动去除特殊标记
• Languages: zh, en, ja, ko, yue (Cantonese)

OpenAI Whisper#

• Model: base (139MB, auto-downloaded)
• Strengths: 90+ languages、翻译模式、多语言场景
• Output: 中文输出繁体字（已知问题，换 small 模型可改善）
• Whisper model sizes:

ModelVRAMSpeedAccuracytiny1GBFastestLowbase1GBFastOKsmall2GBMediumGoodmedium5GBSlowBetterlarge~10GBSlowestBest

Examples#

bash

# Chinese audio → FunASR (default, best quality)
{baseDir}/scripts/asr.py --input meeting.mp3

# Force Chinese language
{baseDir}/scripts/asr.py --input podcast.wav --language zh

# Multilingual audio → Whisper
{baseDir}/scripts/asr.py --input mixed.wav --engine whisper

# Whisper with better model
{baseDir}/scripts/asr.py --input lecture.mp3 --engine whisper --model small

# Translate Chinese speech to English text
{baseDir}/scripts/asr.py --input speech.mp3 --engine whisper --language zh --task translate

# Save transcript to file
{baseDir}/scripts/asr.py --input audio.wav --output transcript.txt

Dependencies#

• funasr + modelscope (FunASR engine)
• openai-whisper (Whisper engine)
• imageio-ffmpeg (bundled ffmpeg binary)
• First run downloads model weights (auto-cached in ~/.cache/)