metadata

license: apache-2.0
language:
  - lv
tags:
  - Whisper
metrics:
  - name: wer
    type: wer
    value: 9.56
pipeline_tag: automatic-speech-recognition

Latvian Whisper small speech recognition model

Trained on combination of:

Common Voice 17, custom selection of all validated clips, max 1000 clips per speaker
Fleurs, test+train+validation

Both regular whisper model and CTranslate2 converted version for use with faster-whisper as part of Home Assistant Whisper integration are available.

For speech generation in Home Assistant use Latvian Piper TTS voice.

To improve speech recognition quality, more data is needed, donate your voice on Balsu talka