jssaluja/nemotron35-punjabi-cer-train-113953-epochs-10-test-1408
This model is a fine-tuned version of nvidia/nemotron-3.5-asr-streaming-0.6b for streaming
Punjabi (Gurbani) automatic speech recognition, prompt-conditioned for target language pa-IN.
It achieves the following results on the validation set (epoch 10, the checkpoint selected and pushed as this model):
- Cer: n/a
- Ncer: n/a
- Ccer: n/a
- Wer: n/a
Training and evaluation data
Training datasets (Sehaj Paath recordings):
- jssaluja/rajinder_singh_corrected
- hustler007/training_bhai_sahib_singh_canada_sehaj_paath
- hustler007/training_bhai_jagtar_singh_jachak_sehaj_paath
- hustler007/training_bhai_gurpreet_singh_nalvi_sehaj_paath
- hustler007/training_bhai_sarwan_singh_sehaj_paath
- hustler007/training_parminder_singh_australia_sehaj_paath_kirtan
- hustler007/training_akashvani_jalandhar_radio_doctor
- hustler007/training_jail_chittian
- hustler007/training_sachi_sakhi_sirdar_kapur_singh
- hustler007/training_rangle_sajjan_akj_audiobook
- hustler007/training_jsk_talks_divine_talk
- hustler007/training_bhai_harpal_singh_zafarnama_katha
- hustler007/training_bhai_simranjeet_singh_tohana_special_talk
- hustler007/training_rangle_sajjan_gurbani_audiobook
- hustler007/training_smtv_podcast_simranjot_makkar
- hustler007/training_bijay_singh
- hustler007/training_khalas_podcast_harsharan_kaur
- hustler007/training_bhai_pinderpal_singh_salok_vaara_te_vadeek
- hustler007/training_satwant_kaur
- hustler007/training_sikh_siyasat_diaspora_representatives
- hustler007/training_sikh_siyasat_gurmeet_singh
- hustler007/training_wikimedia_punjabi_audiobooks
- hustler007/training_hum_hindu_nahin_kahn_singh_nabha
- hustler007/training_heer_waris_shah
- hustler007/training_bhai_pinderpal_singh_jaap_sahib
- hustler007/training_sarbat_media_podcast
- hustler007/training_sikh_siyasat_human_rights
- hustler007/training_smtv_round_table
- hustler007/training_discover_punjab_exclusive_stories
- hustler007/training_bhai_sewa_singh_tarmala_akath_katha
- hustler007/training_jhuthe_police_mukable_di_dastan
- hustler007/training_gurmat_films_guru_nanak_sakhiyan
- hustler007/training_maskeen_ji_full_katha
- hustler007/training_punjabi_chapters_khalsa_raj
- hustler007/training_mehmi_movies_punjabi_short_films
- hustler007/training_tuhituhi_pinderpal_singh_katha
- hustler007/training_gurmat_elite_discover_sikhism
- hustler007/training_gyandasagar_shorts
- hustler007/training_chaupal_chal_mera_putt_3
- hustler007/training_chaupal_pardhan
- hustler007/training_lokdhun_punjabi_latest_movies
- hustler007/training_rajkaregakhalsa_shorts
- hustler007/training_jaanbaaz_rakha_ar_darshi
- hustler007/training_safar_with_nimrat_kaur
- hustler007/training_sikh_siyasat_dr_kanwaljit_singh
Validation datasets (Paath recordings):
- jssaluja/paath_anand_sahib
- jssaluja/paath_anand_sahib_2
- jssaluja/paath_barah_mah_tukhari
- jssaluja/paath_bavan_akhree
- jssaluja/paath_dakhanee_oankaar
- jssaluja/paath_jaitsree_ki_vaar
- jssaluja/paath_japji_sahib
- jssaluja/paath_salok_mehla_9
- jssaluja/paath_siddh_gosst
- jssaluja/paath_sukhmani_sahib
Tokenizer vocabulary was trained on every verse of Guru Granth Sahib.
Training procedure
Training hyperparameters
- Base model:
nvidia/nemotron-3.5-asr-streaming-0.6b - Target language / prompt:
pa-IN - Epochs:
10 - Learning rate:
0.25 - Batch duration (s):
200 - Tokenizer: SentencePiece BPE, vocab size
1024 - Train / validation clip counts:
113953/1408 - Precision:
bf16-mixed - Run type:
full training run
Training results
| Epoch | Train Loss | Val Loss | Cer | Ncer | Ccer | Wer |
|---|---|---|---|---|---|---|
| 1 | 60.4070 | 191.0779 | n/a | n/a | n/a | n/a |
| 2 | 18.5942 | 79.5239 | n/a | n/a | n/a | n/a |
| 3 | 13.0003 | 55.9048 | n/a | n/a | n/a | n/a |
| 4 | 8.9209 | 49.8107 | n/a | n/a | n/a | n/a |
| 5 | 8.5751 | 45.3670 | n/a | n/a | n/a | n/a |
| 6 | 8.0170 | 44.1199 | n/a | n/a | n/a | n/a |
| 7 | 5.2882 | 44.1397 | n/a | n/a | n/a | n/a |
| 8 | 3.8503 | 44.5815 | n/a | n/a | n/a | n/a |
| 9 | 5.0727 | 44.4870 | n/a | n/a | n/a | n/a |
| 10 | 9.3243 | 43.8441 | n/a | n/a | n/a | n/a |
Framework versions
- NeMo: git+https://github.com/NVIDIA-NeMo/NeMo.git@main
- PyTorch: 2.13.0+cu130