README.md

metadata

license: cc-by-nc-4.0
language:
  - hyw
datasets:
  - mozilla-foundation/common_voice_16_1
  - google/fleurs
  - ReRooted
pipeline_tag: automatic-speech-recognition
tags:
  - audio-to-audio
  - text-to-speech
  - seamless_communication

SeamlessM4T v2 ASR for Western Armenian

This model is a fine-tuned version of the facebook/seamless-m4t-v2-large. Initially, it was fine-tuned on the Common Voice 16.1 and Google Fleurs datasets. Subsequently, it was further fine-tuned on the ReRooted corpus. The model achieves the following results on the test sets:

• CV_wer: 0.308
• CV_cer: 0.07
• GF_wer: 0.311
• GF_cer: 0.094

After fine-tuning on Western Armenian data, the model occasionally translates Eastern Armenian speech into Western Armenian.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-6
• train_batch_size: 4
• eval_batch_size: 1
• seed: 43
• optimizer: Adam with betas=(0.9, 0.98) and epsilon=1e-08
• lr_scheduler_type: MyleLR
• lr_scheduler_warmup_steps: 100

Framework versions

• Pytorch 2.1.1
• fairseq2==0.2.0