bangla-asr

Bangla ASR — Fine-tuned Whisper-medium for Bengali Speech Recognition

Released by Poseidon / PSDN AI.

This repository contains a full-model fine-tuned Whisper-medium checkpoint for Bengali automatic speech recognition. It is intended for Bengali ASR research and inference.

This is the checkpoint corresponding to the model reported in our Bangla-10K paper.

• Experiment identifier: V7-E0-fullft-seed42
• Checkpoint: checkpoint-4000
• Architecture: Whisper-medium (WhisperForConditionalGeneration)

The released config.json records the immediate trainer initialization path as a merged V6 checkpoint (/v7_vol/base/v6e0a_merged). That merged checkpoint’s own _name_or_path, and the run configuration for this experiment, both identify the Hugging Face base as bengaliAI/tugstugi_bengaliai-asr_whisper-medium.

Inference

from transformers import WhisperForConditionalGeneration, WhisperProcessor

model_id = "psdn-ai/bangla-asr"
processor = WhisperProcessor.from_pretrained(model_id)
model = WhisperForConditionalGeneration.from_pretrained(model_id)

A typical decode for a 16 kHz mono waveform:

inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
predicted_ids = model.generate(inputs.input_features, language="bn", task="transcribe")
text = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]

The saved generation config already sets language="bn" and task="transcribe".

Notes

Benchmark numbers, dataset licenses, training hours, and other paper details will be added to this model card later. Please refer to the Bangla-10K paper for the full evaluation protocol.

Optimizer, scheduler, and RNG trainer state are not included in this release.