GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
OmiKhan
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
OmiKhan
Model Tree
bengaliAI/tugstugi_bengaliai-regional-asr_whisper-medium
Input Modalities
Output Modalities
Supported Functionality
Word error rate (%), greedy decoding, one normalizer for every system (bangla_asr.common.text.normalize_bangla_text).
In brackets: the spelling-aware WER, which does not count differences that human annotators confirmed as spelling only
(ছিল / ছিলো).
| Test set | This model | Zero-shot base model |
|---|---|---|
| DL Sprint 4.0 eval (media speech, 4,234 clips) | 11.18 (8.82) | 22.70 (15.82) |
| FLEURS test | 15.60 (10.89) | 21.54 (14.56) |
| Common Voice test | 10.13 (6.61) | 14.80 (10.15) |
| Kathbath validation | 10.28 (5.87) | 14.81 (9.02) |
| Average | 11.80 (8.04) | 18.46 (12.39) |
Against the open Bangla models scored with the same harness it is best on DL Sprint, Common Voice and Kathbath among the models not trained on those sets; on FLEURS the non-commercial Hishab conformer is ahead (14.24). Against 605 human-verified references the order of the systems is unchanged. Whole raw recordings (22, 20.4 h, cut by voice detection): 14.2% WER. Speed: real-time factor about 0.04 on a T4.
python
from transformers import pipeline asr = pipeline("automatic-speech-recognition", model="OmiKhan/bangla-asr-whisper-medium")print(asr("clip.wav", generate_kwargs={"language": "bn", "task": "transcribe"})["text"])Audio is resampled to 16 kHz. The model was trained and scored on clips of 1–25 s; for long recordings, cut the audio
at pauses (for example with Silero VAD) or pass chunk_length_s=30.
Full fine-tuning, 2 epochs, learning rate 1e-5, from the tugstugi Whisper-medium checkpoint. The released weights are a weight average of two runs with the same recipe, chosen by a rule fixed before any test score (lowest average WER over three development sets): 0.25 × M1 (DL Sprint, 60 h) + 0.75 × M5-300 (all data below, ~300 h).
| Data | Hours used | License |
|---|---|---|
| DL Sprint 4.0 long-form ASR (Bengali-Loop), training recordings, forced-aligned | ~60 | no license stated; see below |
| Common Voice 27.0, Bengali (train) | ~34 | CC0 |
| Kathbath, Bengali (train) | ~34 | CC BY 4.0 |
| Shrutilipi, Bengali | ~28 | CC BY 4.0 |
| OpenSLR 53, Large Bengali ASR training data | ~144 | CC BY-SA 4.0 |
The test sets are the corpora's official test (or validation) splits; OpenSLR 53 sentences that also occur in a test set were removed before training.
The DL Sprint 4.0 audio comes from public Bangla YouTube channels and is subject to the competition's rules, which state no data license; no audio is included in this repository. If you are a rights holder or the organizers and want a model or derived material removed, open an issue at github.com/omikhan4901/asr.
Mehboob Ehsan Khan, Nabigah Bin Sayeed, Nafisa Rodela. Why N-best Rescoring Plateaus for Bangla Speech Recognition, and How Far Fixing the Cause Goes. CSE499 capstone, North South University, 2026. github.com/omikhan4901/asr
from transformers import pipeline asr = pipeline("automatic-speech-recognition", model="OmiKhan/bangla-asr-whisper-medium")print(asr("clip.wav", generate_kwargs={"language": "bn", "task": "transcribe"})["text"])