Shenava 1.0: Open Streaming Persian ASR and Captioning
Collection
Public Shenava 1.0 release: Persian ASR models, Phase A/B datasets, AL corrections, eval artifacts, and demos. • 60 items • Updated
How to use Reza2kn/visualears-fastconformer-fa-69m-256x40-warmstart with NeMo:
import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.ASRModel.from_pretrained("Reza2kn/visualears-fastconformer-fa-69m-256x40-warmstart")
transcriptions = asr_model.transcribe(["file.wav"])This repository is auto-updated from the live VisualEars 69M warm-start training run.
/workspace/train69_runs/fa69m_256x40_warmstart_full9669_20260615T015714Z/fa69m_256x40_warmstart_bpe1024/2026-06-15_01-57-22/checkpoints/fa69m_256x40_warmstart_bpe1024--val_wer=0.1551-epoch=0.ckptcheckpoints/fa69m_256x40_warmstart_bpe1024--val_wer=0.1551-epoch=0.ckpt0.1551837980369 bytes5cca73aee0f42d9e577eb4c9421a3d8c4e09777a55d18264cbd7553ab44013542026-06-15T16:21:08Zd_model: 256n_heads: 4n_layers: 40pred_hidden: 640joint_hidden: 64069.627906M/workspace/train32_runs/fa32m_streaming_bpe1024_full9669_20260614T130932Z/fa32m_streaming_bpe1024/2026-06-14_13-09-40/checkpoints/fa32m_streaming_bpe1024--val_wer=0.1720-epoch=0.ckpt61931648226[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15]0New encoder layers are near-identity initialized; see launch_stats.json / best_checkpoint.json for details.
/workspace/train69_runs/fa69m_256x40_warmstart_full9669_20260615T015714Z/fa69m_256x40_warmstart_bpe1024_final.nemofinal/fa69m_256x40_warmstart_bpe1024_final.nemo279511040 bytesc6cbaf40d12f635024be0c748fe9b4613604bd4decc31289ff66c55adbda6ea0