RaivisDejus/latvian-deepspeech

Latvian DeepSpeech speech recognition model

For more information about the DeepSpeech see https://deepspeech.readthedocs.io/en/r0.9/

This model has two versions:

• The regular (output_graph.*) - This is more precise but model files take more space
• The lite (output_graph-lite.*) - This is a bit less precise, but model files are smaller

Language model (scorer) has three versions:

• The regular (kenlm-lv.scorer) - The biggest and most precise, built with wikipedia, tilde and europarl data
• The wiki (kenlm-lv-wiki.scorer) - Smaller, but still quite comprehensive, built with wikipedia data
• The wiki (kenlm-lv-lite.scorer) - Smallest, built with subset of wikipedia data

Model is trained on:

• https://huggingface.co/datasets/mozilla-foundation/common_voice_12_0
• https://huggingface.co/datasets/google/fleurs

Language model (scorer) is generated from:

• https://huggingface.co/datasets/RaivisDejus/latvian-text

Recognition precision

Regular model

Regular model with FULL scorer

• CV 12 only - WER: 0.150664, CER: 0.044672, loss: 6.493098
• CV 12 + Fleurs - WER: 0.296679, CER: 0.109360, loss: 85.185188

Regular model with WIKI scorer

• CV 12 only - WER: 0.183172, CER: 0.053100, loss: 6.493098
• CV 12 + Fleurs - WER: 0.321415, CER: 0.114797, loss: 85.185188

Regular model with LITE scorer

• CV 12 only - WER: 0.230378, CER: 0.064393, loss: 6.493098
• CV 12 + Fleurs - WER: 0.335594, CER: 0.119691, loss: 85.185188

Lite model

Lite model with FULL scorer

• CV 12 only - WER: 0.390653, CER: 0.174240, loss: 18.772360
• CV 12 + Fleurs - WER: 0.570163, CER: 0.288619, loss: 147.965317

Lite model with WIKI scorer

• CV 12 only - WER: 0.429002, CER: 0.186182, loss: 18.772360
• CV 12 + Fleurs - WER: 0.588117, CER: 0.293284, loss: 147.965317

Lite model with LITE scorer

• CV 12 only - WER: 0.484594, CER: 0.203621, loss: 18.772360
• CV 12 + Fleurs - WER: 0.595669, CER: 0.293980, loss: 147.965317

Usage

This DeepSpeech model can be used to add voice commands to different systems, like Home Assistant with help of Rhasspy Voice Assistant.

Demo is available https://github.com/raivisdejus/DeepSpeechLV