https://github.com/shuyib/zindi_mcv_swahilli

How I used Seamless m4t large to get to the top 5 of the mozilla common voice competition hosted on Zindi
https://github.com/shuyib/zindi_mcv_swahilli

asr-model hackathon mozilla-common-voice seamlessm4t stt swahili voice-recognition zindi-hackathon

Last synced: 24 days ago
JSON representation

How I used Seamless m4t large to get to the top 5 of the mozilla common voice competition hosted on Zindi

• Host: GitHub
• URL: https://github.com/shuyib/zindi_mcv_swahilli
• Owner: Shuyib
• License: apache-2.0
• Created: 2023-12-12T06:55:43.000Z (about 1 year ago)
• Default Branch: main
• Last Pushed: 2023-12-27T10:56:33.000Z (12 months ago)
• Last Synced: 2024-04-16T03:50:55.078Z (8 months ago)
• Topics: asr-model, hackathon, mozilla-common-voice, seamlessm4t, stt, swahili, voice-recognition, zindi-hackathon
• Language: Python
• Homepage:
• Size: 14.6 KB
• Stars: 1
• Watchers: 1
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# zindi_mcv_swahilli

public word error rate: **0.114294524**   

private word error rate: **0.112809661**   

How I used Seamless m4t large to get to the top 5 of the [mozilla common voice competition](https://zindi.africa/competitions/mozilla-foundation-mozilla-common-voice-hackathon-i-nairobi). I only downloaded the `test.tar.gz` directory later I unzipped it and resampled all the audio to 16KHz. I noticed that there was some audio that was muffled, and was pretty bad as is due to the sampling rates that were set. Anyways, the script I used to do the conversion is called `prepare_files.sh`. Follow the instructions to install [seamless m4t large](https://github.com/facebookresearch/seamless_communication). I performed inference on each audio file `python asr.py` the output was then saved to **asr_results.csv** then it was formatted to a certain format needed for Zindi with `python clean_submission.py`. 

## You can do all this in one step

```bash

make run

```

## Lesson

Review huggingface leaderboard for the ASR models. Look for one with the fastest and the most accurate. 

[leaderboard](https://huggingface.co/models?other=hf-asr-leaderboard)

Facebook/meta have a lot of Speech to text models. Look for one that is capable of doing Speech to text. The ones that primarily do one thing seem to be the best.