LoRA_bengaliAI_tugstugi_bengaliai-asr_whisper-medium
Fine-tuned Bengali ASR model for long-form, noisy, real-world audio, built on top of
bengaliAI/tugstugi_bengaliai-asr_whisper-medium.
This model is adapted for long-form Bangla speech transcription using parameter-efficient fine-tuning (LoRA) and was developed for challenging multi-speaker Bengali media such as dramas, broadcasts, and other in-the-wild recordings.
Model Description
- Base model:
bengaliAI/tugstugi_bengaliai-asr_whisper-medium
- Task: Automatic Speech Recognition (ASR)
- Language: Bengali / Bangla
- Adaptation: LoRA-based fine-tuning for long-form Bengali ASR
- Target setting: Noisy, continuous, real-world Bengali audio
Training Data
This model was fine-tuned on the Bengali-Loop long-form Bengali ASR corpus.
The training pipeline first converts long recordings into Whisper-compatible segments, then fine-tunes the base model on those aligned chunks.
Long-Form Inference Pipeline
For best results on long audio, use this model with a long-form inference pipeline that includes:
- Voice Activity Detection (VAD)
- Context-aware chunking / segmentation
- Optional denoising or source separation for noisy media
- Bengali-specific text normalization / cleanup
This model is intended for long-form transcription workflows, not only short isolated utterances.
Performance
On the Bengali-Loop private test setting reported in our paper:
This improves substantially over the untuned long-form baseline built on the same base model.
Intended Use
Suitable for:
- Long-form Bengali transcription
- YouTube / podcast / drama transcription
- Broadcast or conversational Bangla audio
- Research on low-resource ASR
- Bengali speech pipeline development
Limitations
Performance may still degrade under:
- Heavy background music or sound effects
- Extreme overlap between speakers
- Strong Bengali-English code-switching
- Highly emotional or distorted speech
- Domain shifts outside Bengali media-style audio
This model is a research model and should be evaluated before production deployment.
Usage
Use with the Hugging Face transformers ASR pipeline or in a custom Whisper inference setup.
For long recordings, chunked decoding is strongly recommended.
Acknowledgements
This model is built on top of the excellent tugstugi_bengaliai-asr_whisper-medium base model from Bengali.AI.
