wscribe 0.1.5

ez audio transcription tool with flexible processing and post-processing options

Table of Contents

1. wscribe
   1. Getting started
      1. Installation
      2. Usage
   2. Numbers
   3. Roadmap
      1. Processing Backends
      2. Transcription Features
      3. Inference interfaces
      4. Audio sources
      5. Distribution
   4. Contributing
      1. Testing

wscribe

Getting started

wscribe is yet another easy to use front-end for whisper specifically for transcription. It aims to be modular so that it can support multiple audio sources, processing backends and inference interfaces. It can run both on CPU and GPU based on the processing backend. Once transcript is generated, editing/correction/visualization of the transcript can be done manually with the wscribe-editor.

It was created at sochara because we have a large volume of audio recordings that need to be transcribed and eventually archived. Another important need was that we needed to verify and manually edit the generated transcript, I could not find any open-source tool that checked all the boxes. Suggested workflow is generating word-level transcript(only supported in json export) and then editing the transcript with the wscribe-editor.

Currently, it supports the following. Check roadmap for upcoming support.

• Processing backend: faster-whisper
• Audio sources: Local files (Audio/Video)
• Inference interfaces: Python CLI
• File exports: JSON, SRT, WebVTT

Installation

These instructions were tested on NixOS:Python3.10 and ArchLinux:Python3.10 but should work for any other OS, if you face any installation issues please feel free to create issues. I’ll try to put out a docker image sometime.

1. 1. Set required env var

   • WSCRIBE_MODELS_DIR : Path to the directory where whisper models should be downloaded to.

     export WSCRIBE_MODELS_DIR=$XDG_DATA_HOME/whisper-models # example

2. 2. Download the models

   3. Recommended

      • Recommended way for downloading the models is to use the helper script, it’ll download the models to WSCRIBE_MODELS_DIR.

        cd /tmp # temporary script, only needed to download the models
        curl https://raw.githubusercontent.com/geekodour/wscribe/main/scripts/fw_dw_hf_wo_lfs.sh
        chmod u+x fw_dw_hf_wo_lfs.sh
        ./fw_dw_hf_wo_lfs.sh tiny # other models: tiny, small, medium and large-v2
        

   4. Manual

      You can download the models directly from here using git lfs, make sure you download/copy them to WSCRIBE_MODELS_DIR

3. 3. Install wscribe

   Assuming you already have a working python>=3.10 setup

   pip install wscribe
   

Usage

# wscribe transcribe [OPTIONS] SOURCE DESTINATION

# cpu
wscribe transcribe audio.mp3 transcription.json
# use gpu
wscribe transcribe video.mp4 transcription.json --gpu
# use gpu, srt format
wscribe transcribe video.mp4 transcription.srt -g -f srt
# use gpu, srt format, tiny model
wscribe transcribe video.mp4 transcription.vtt -g -f vtt -m tiny
wscribe transcribe --help # all help info

Numbers

• These numbers are from machine under normal web-browsing workload running on a single RTX3050
• Audio conversion takes around 1s
• Also check this explanation about the speed difference.

device	quant	model	original playback	transcription	playback/transcription
cuda	float16	tiny	6.3m	0.1m	68x
cuda	float16	small	6.3m	0.2m	29x
cuda	float16	medium	6.3m	0.4m	14x
cuda	float16	large-v2	6.3m	0.8m	7x
cpu	int8	tiny	6.3m	0.2m	25x
cpu	int8	small	6.3m	1.3m	4x
cpu	int8	medium	6.3m	3.6m	~1.7x
cpu	int8	large-v2	6.3m	3.6m	~0.9x

Roadmap

Processing Backends

• faster-whisper
• whisper.cpp
• WhisperX, diarization is something to look forward to

Transcription Features

• Add support for diarization
• Add translation
• Add VAD/other de-noising stuff etc.
• Add local llm integration with llama.cpp or something similar for summary and othe possible things. It can be also used to generate more accurate transcript. Whisper mostly generates sort of a subtitle, for converting subtitle into transcription we need to group the subtitle. This can be done in various ways. Eg. By speaker if diarization is supported, by time chunks etc. By using LLMs or maybe other NLP techniques we’ll also be able to do this with things like break in dialogue etc. Have to explore.

Inference interfaces

• [-] Python CLI
  • Basic CLI
  • Improve summary statistics
• REST endpoint
  • Basic server to run wscribe via an API.
  • Possibly add glue code to expose it via CFtunnels or something similar

Audio sources

• Local files
• Youtube
• Google drive

Distribution

• Python packaging
• Docker/Podman
• Package for Nix
• Package for Arch(AUR)

Contributing

All contribution happens through PRs, any contributions is greatly appreciated, bugfixes are welcome, features are welcome, tests are welcome, suggestions & criticism are welcome.

Testing

• make test
• See other helper commands in Makefile