surindersinghssj/surt-small-v1-kirtan

Surt Small v1 Kirtan — Gurbani Kirtan ASR

Fine-tuned from surt-small-v1 (Sehaj Path model) on kirtan audio data for Gurbani kirtan transcription and forced alignment.

Model Details

Parameter	Value
Base model	surindersinghssj/surt-small-v1-training (step 3400)
Language	Punjabi (Gurmukhi script)
Task	Transcribe
WER	32.65%
CER	24.62%
Training data	260 kirtan samples from 11 artists

Training

• Dataset: surindersinghssj/gurbani-asr-whisper-aligned — 260 train / 31 eval
• Artists: 11 kirtan artists (Bhai Manpreet Singh Kanpuri, Bhai Anantvir Singh, etc.)
• Hardware: NVIDIA A40, single GPU
• Training time: ~21 minutes (500 steps)
• Effective batch size: 16 (batch 8 x gradient accumulation 2)
• Learning rate: 2e-5 (lower than base training since continuing from fine-tuned model)
• Scheduler: Cosine decay with 50 warmup steps

WER Progression

Step	Epoch	WER	CER
0 (pre-training)	—	118.19%	92.82%
100	3.0	61.63%	48.85%
200	8.9	41.63%	30.34%
300	14.7	31.84%	23.19%
350	17.7	31.43%	23.19%
500 (final)	29.4	32.65%	24.62%

Related Models

Model	Use case	WER
surt-small-v1	Sehaj Path transcription	14.88%
surt-small-v1-kirtan (this)	Kirtan transcription/alignment	32.65%

Usage

from transformers import WhisperProcessor, WhisperForConditionalGeneration

processor = WhisperProcessor.from_pretrained("openai/whisper-small")
model = WhisperForConditionalGeneration.from_pretrained("surindersinghssj/surt-small-v1-kirtan")

import librosa
audio, sr = librosa.load("kirtan_audio.wav", sr=16000)
input_features = processor(audio, sampling_rate=16000, return_tensors="pt").input_features
predicted_ids = model.generate(input_features)
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
print(transcription)

Note: Use WhisperProcessor.from_pretrained("openai/whisper-small") for the processor.

Limitations

• Trained on only 260 samples — more data would significantly improve performance
• Best WER was at step 350 (31.43%), slight overfitting after that
• Audio-text alignment in training data is imperfect (kirtan involves repetition and musical phrasing)

License

Apache 2.0