baniguru_v1

BaniGuru Gurbani ASR — Offline

BaniGuru Gurbani ASR is a Punjabi (Gurmukhi) automatic speech recognition model fine-tuned specifically for Gurbani speech.

The model is designed for transcription of both Gurbani Kirtan and spoken/recited Gurbani such as Sehaj Paath.

It is based on AI4Bharat's IndicConformer hybrid CTC-RNNT architecture and has been fine-tuned on approximately 150 hours of Gurbani speech.

Model Details

Model Description

• Model type: Automatic Speech Recognition (ASR)
• Architecture: IndicConformer Hybrid CTC-RNNT
• Parameters: ~120M
• Language: Punjabi (pa)
• Script: Gurmukhi
• Domain: Gurbani
• Mode: Offline / non-streaming
• Framework: NVIDIA NeMo / AI4Bharat NeMo
• License: MIT
• Fine-tuned from: ai4bharat/indicconformer_stt_pa_hybrid_ctc_rnnt_large
• Training dataset: sukhpalsingh/speech-data-v1

The model retains the multilingual tokenizer/vocabulary structure of the underlying AI4Bharat model while being fine-tuned specifically for Punjabi Gurbani recognition.

Uses

Direct Use

The model can be used to transcribe Punjabi Gurbani audio written in the Gurmukhi script.

It is intended primarily for:

• Gurbani Kirtan transcription
• Sehaj Paath / Gurbani recitation transcription
• Punjabi religious speech recognition
• Gurbani audio indexing and search
• Generating draft transcripts for Gurbani recordings

Downstream Use

The model can be integrated into applications such as:

• Gurbani transcription services
• Searchable Kirtan archives
• Gurbani subtitle generation
• Audio indexing systems
• Gurbani applications requiring speech-to-text

Limitations

The model has been specifically fine-tuned for Gurbani and should not be assumed to provide the same accuracy for general Punjabi conversational speech.

Kirtan is substantially more difficult to recognize than clear recited speech because of singing, musical accompaniment, variable tempo, repetition, and recording conditions.

Accuracy can therefore vary considerably depending on:

• singing style
• background instruments
• recording quality
• microphone quality
• overlapping voices
• pronunciation and vocal style
• audio duration and segmentation

The model's output should not be assumed to be an authoritative rendering of Gurbani text without appropriate verification.

Training Details

Training Data

The model was fine-tuned using approximately 150 hours of Gurbani audio.

Training data includes Gurbani speech with emphasis on domains such as Kirtan and recited Gurbani.

Dataset:

sukhpalsingh/speech-data-v1

Preprocessing

Training samples longer than 30 seconds were excluded from the training manifest.

Audio/text pairs were prepared for Punjabi (pa) recognition using Gurmukhi transcripts.

Training Hyperparameters

• Batch size: 8
• Gradient accumulation: 1
• Learning rate: 1e-5
• Maximum epochs: 50
• Maximum training audio duration: 30 seconds
• Language ID: pa

The model was fine-tuned while preserving the pretrained multilingual tokenizer and vocabulary structure.

Evaluation

The model was evaluated separately on two Gurbani domains:

1. Kirtan — sung Gurbani with the additional difficulty of music, melody and variable vocal delivery.
2. Sehaj Paath — clearer recited Gurbani speech.

Metrics

Evaluation uses:

• WER — Word Error Rate
• CER — Character Error Rate

CER is particularly useful for evaluating Gurmukhi transcription because it measures recognition errors at the character level.

Results

These results reflect different acoustic conditions and should not be interpreted as equivalent test sets. Kirtan represents a substantially more difficult recognition environment than clear Gurbani recitation.

Technical Specifications

Model Architecture and Objective

The model is based on AI4Bharat's IndicConformer Hybrid CTC-RNNT architecture.

It contains approximately 120 million parameters and supports both CTC and RNNT objectives/decoding through the hybrid architecture.

The fine-tuned model preserves the underlying multilingual tokenizer configuration while targeting Punjabi (pa) Gurbani recognition.

Software

The model was trained using:

• PyTorch
• NVIDIA NeMo
• AI4Bharat NeMo v2

Acknowledgements

This model is fine-tuned from AI4Bharat's IndicConformer Punjabi ASR model.

Base model:

ai4bharat/indicconformer_stt_pa_hybrid_ctc_rnnt_large

Training dataset:

sukhpalsingh/speech-data-v1