riasatsk/Bengali_ASR

bengali-asr — 1:1 Verbatim ONNX (rebranded folder, ai4bharat weights verbatim)

Folder bengali-asr/ is separate from neo/ (BanglaNeo bn-only) and models/ original .nemo. Display name bengali-asr but ONNX metadata keeps original_source: ai4bharat/indicconformer_stt_bn_hybrid_ctc_rnnt_large for audit.

• Verbatim: Conformer-L 120M 17×512 encoder + preprocessor AudioToMel 80 inside ONNX, Conv1d 512→5633 (22*256+blank 5632) kept, 22 tokenizer triples verbatim (tokenizer/ 66 files). No slicing like neo 257. Input wav [1,T] 16k mono float32 + wav_len [1] → logits [1,T/4,5633] → slice bn [256:512]+blank 5632 → 257.
• Files:
  • models/bengali-asr-wav-ctc-5633.onnx (+.data 469 MB) — fused wav→5633, 472 MB, primary 1:1
  • models/bengali-asr-encoder-wav.onnx (+.data 458 MB) — wav→enc 512
  • models/bengali-asr-decoder-ctc-5633.onnx (+.data 11 MB) — enc→5633
  • models/tokenizer/ 22 langs hash names verbatim
  • models/config.json (display_name: bengali-asr, original_source: ai4bharat/..., vocab 5632/5633, langs 22)
  • models/tokenizer_map.json 22 entries

Why not BanglaNeo: BanglaNeo neo/models/bangla-neo-conformer-bn-ctc.onnx 462 MB is bn-only 257, mel outside, rebranded metadata (no ai4bharat). bengali-asr is ~5-10% slower (5633 argmax) but keeps 22 langs + wav in + audit provenance. Use BanglaNeo for small/fast bn only, bengali-asr for verbatim audit.

Quick start (new identity, separate folder, port 8002):

# Reuse .venv311 (Python 3.11, onnxruntime CPU, no nemo at inference)
C:\Users\Riasat\BengaliSTT\.venv311\Scripts\python.exe bengali-asr/src/transcribe_bengali_asr.py --audio neo/test_bn.mp3 --lang bn
# -> bengali-asr (bn): আমি বাংলায় কথা বলছি  logits(1,62,5633) sliced [256:512]+5632

# Web 1:1 verbatim
C:\Users\Riasat\BengaliSTT\.venv311\Scripts\python.exe bengali-asr/src/app_bengali_asr.py
# http://localhost:8002  (vs neo 8001, old app 8000)
# curl -F audio=@sample.wav -F lang=bn http://localhost:8002/transcribe

Compare:

• BanglaNeo 8001: bn 257, mel in, 462 MB, http://localhost:8001 — small/fast
• bengali-asr 8002: 22 5633, wav in, 472 MB, verbatim — audit

Providers: CUDA→CPU auto pick_providers() onnxruntime CPUExecutionProvider default, onnxruntime-gpu for CUDA.

Note: Folder bengali-asr/ is separate, does not overwrite neo/ or models/. Both can run together (8001 + 8002).