stt_bn_fastconformer_ctc

Bengali speech-to-text, FastConformer-CTC large (115.6 M parameters, 8× subsampling), greedy CTC decode, no language model, 1,024-piece Bengali BPE, trained on 1,692 hours of human-supervised Bengali.

Weights: stt_bn_fastconformer_ctc.nemo, SHA-256 bea300ce4f1f1df1fa4707562ddaa7c64db6f01121ef3d64a6333635f3626572.

Output

Input 16 kHz mono WAV/FLAC; output normalised Bengali text without punctuation.

# pip install -q 'nemo_toolkit[asr]==2.7.3'
import nemo.collections.asr as nemo_asr
m = nemo_asr.models.ASRModel.from_pretrained("ehzawad/stt_bn_fastconformer_ctc")
print(m.transcribe(["your_audio_16k_mono.wav"])[0].text)

Dataset

1,084,696 clips, 1,692.4 hours, all human-read or human-transcribed; no pseudo-labels, no synthetic audio.

Source	Hours	Licence
Bengali.AI Speech (Kaggle bengaliai-speech, train split, filtered by tugstugi's WER < 15 id list; transcripts are the competition's own)	726.5	CC-BY-4.0
IndicVoices (Bengali)	574.4	CC-BY-4.0
OpenSLR-53	159.0	CC-BY-SA-4.0
SUBAK.KO	105.8	CC-BY-4.0
Kathbath (Bengali)	81.5	CC-BY-4.0
Common Voice (validated)	39.5	CC0
FLEURS bn_in train split	5.8	CC-BY-4.0

Base checkpoint: nvidia/stt_en_fastconformer_ctc_large, © NVIDIA, CC-BY-4.0. Full credits, licence notices and modifications: SUPPLEMENT.md.

Benchmark

FLEURS Bengali official test split, 920 recordings, corpus WER on punctuation-normalised text, greedy decoding, no language model: WER 17.12 %, CER 5.12 %.

Under the same scorer on the same 920 recordings, the public comparison in bengali-asr-benchmark: hishab Conformer Large 14.37 %, Whisper Medium (tugstugi) 16.23 %, this model 17.12 %, wav2vec2 20.64 %, hishab FastConformer 36.22 %. On the official test splits of corpora none of these models trained on, this model leads: Vaani test 24.66 %, SPRING-INX R1 39.24 %, R2 38.41 % (1,000-utterance subsets); on Kathbath test it reads 13.4 % against Conformer Large's 9.3 %; on IndicVoices conversational 12.6 % (it trained on that corpus's train split; the others read 25–33 %).

Held-out panels, the benchmark tables and the limitations: SUPPLEMENT.md.

Licence

Weights: CC-BY-SA-4.0. Attribution and share-alike follow from OpenSLR-53; the required notices are in SUPPLEMENT.md.

Citation

@misc{zawad2026bengalifastconformerctc,
  title        = {Bengali FastConformer-CTC: Expanded-Corpus Speech Recognition Under a Registered Protocol},
  author       = {Zawad, Emrul},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/ehzawad/stt_bn_fastconformer_ctc}}
}