Skip to main content

FASR: Fast Automatic Speech Recognition

Project description

🪐 项目: fasr

Fast Auto Speech Recognition

📋 简介

fasr是一款快速且易于使用的python库,它源于FunASR,专注于推理性能,目标是成为一个工业级别的python语音识别推理库。

📋 安装

fasr可以通过直接通过pip安装,但是如果需要使用gpu,需要手动安装pytorch和onnxruntime-gpu

  • 安装pytorch: 通过官网安装对应cuda版本
  • 安装onnxruntime-gpu: 通过官网安装对应cuda版本
  • 安装fasr
pip install fasr --upgrade

📋 使用

  • 下载模型
fasr prepare online # 实时语音识别模型
fasr prepare offline # 离线语音识别模型
  • 构建离线语音识别pipeline
from fasr import AudioPipeline

# 语音识别pipeline: 端点检测->语音识别->添加标点
asr = AudioPipeline().add_pipe('detector').add_pipe('recognizer').add_pipe('sentencizer')

# 运行pipeline并获取文本
url = "https://xxxxx.mp3"
audio = asr(url)
for channel in audio.channels:
    print(channel.text)

# 批次处理
urls = get_urls()
audios = asr.run(urls)
# 获取文本
for audio in audios:
    for channel in audio.channels:
        print(channel.text)
  • 构建实时语音识别pipeline
from fasr import AudioPipeline, Audio
asr = AudioPipeline().add_pipe("online_detector").add_pipe("online_recognizer")

# 模拟流式接收音频流
waveform, sample_rate, is_last = get_waveform_from_stream()
audio_chunk = Audio().append_channel()
audio_chunk.is_last = ia_last
audio_chunk.channels[0].waveform = waveform
audio_chunk.channels[0].sample_rate = sample_rate
for token in asr(audio_chunk).channels[0].stream:
    print(token.text)
# 清除cache,恢复组件初始状态
asr.get_pipe("online_recognizer").component.reset()
asr.get_pipe("online_detector").component.reset()

📋 性能对比

双通道音频

vad->asr->punc

测试结果

cpu: Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz (wpai cpu 2核)

gpu: rtx6000 (wpai vgpu 20)

框架 耗时 推理速度 加速比
funasr 368.8s 46.34 1.0
fasr 153.92s 111.03 2.4

vad

测试结果

cpu: Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz (wpai cpu 2核)

gpu: rtx6000 (wpai vgpu 20)

框架 耗时 推理速度 加速比
funasr 219.8s 77.75 1.0
fasr 86.32s 197.98 2.55

单通道音频

pipeline

vad->asr->punc

测试结果

cpu: Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz (wpai cpu 2核)

gpu: rtx6000 (wpai vgpu 20)

框架 耗时 推理速度 加速比
funasr 123.8s 22.05 1.0
fasr 59.04s 46.24 2.1

vad

测试结果

cpu: Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz (wpai cpu 2核)

gpu: rtx6000 (wpai vgpu 20)

框架 耗时 推理速度 加速比
funasr 59.26s 46.07 1.0
fasr 36.84s 74.1 1.61

AISHELL

pipeline

vad->asr->punc

测试结果

cpu: Intel(R) Xeon(R) Silver 4210 CPU @ 2.20GHz (wpai cpu 2核)

gpu: rtx6000 (wpai vgpu 20)

框架 耗时 推理速度 加速比
funasr 123.8s 18.65 1.0
fasr 59.04s 32.71 1.8

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

fasr-0.2.8.tar.gz (5.7 MB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

fasr-0.2.8-py3-none-any.whl (5.8 MB view details)

Uploaded Python 3

File details

Details for the file fasr-0.2.8.tar.gz.

File metadata

  • Download URL: fasr-0.2.8.tar.gz
  • Upload date:
  • Size: 5.7 MB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: uv/0.6.4

File hashes

Hashes for fasr-0.2.8.tar.gz
Algorithm Hash digest
SHA256 f8d069c451e3c28253cdee569bf2cfced0ad166d7bbc46dc5fb197219c73e74c
MD5 3ebe6a58d18137efc35fa1de8e86f0b1
BLAKE2b-256 d0ac78785212aa80f53d5a7b0893921c329111d8e2c6424cf4838b0d7f1e15e3

See more details on using hashes here.

File details

Details for the file fasr-0.2.8-py3-none-any.whl.

File metadata

  • Download URL: fasr-0.2.8-py3-none-any.whl
  • Upload date:
  • Size: 5.8 MB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: uv/0.6.4

File hashes

Hashes for fasr-0.2.8-py3-none-any.whl
Algorithm Hash digest
SHA256 e0a3519826112e1b1f4141c498d26569a2260f76c373d0f097959ab2bc229c37
MD5 8c48083c7808fde03025feb0add870c7
BLAKE2b-256 8e2814e0cfe406013e7410663408a63c694e0d53d1788debac86e0ae47ff1012

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page