Developer lab

Hold to speak — ONNX vs PyTorch transcribe, export, sound match, enrollment.

1. Checkpoint → ONNX

Copy to Android assets Verify with audio below

Writes to web/exports/onnx_mobile/student_ctc.onnx

PyTorch needs a .ckpt file — not included in android_handoff. Copy from the training server into ManipuriAsr/checkpoints/ (see checkpoints/README.md).

2. Transcribe

Shop ID (voice profile)

Choose audio

Hold the button to speak, or choose an audio file

ONNX

—

PyTorch checkpoint

—

3. Sound match (no text / no ASR)

Find a short entity sound (e.g. onion) inside your clip, or match the whole recording. References in web/kirana/catalog.json.

Match mode Scan all catalog items

Match file

—

4. Finetune samples (phrase experiment)

Teaches specific words, not general voice. For shopkeepers use §5 instead.

Record ~10 short phrases for a personal finetune of last.ckpt. Word list: kirana_natural_word_order_500.json.

Loading…

After recording:
conda activate manipuri-asr
bash scripts/run_finetune.sh

5. Shop voice enrollment

Read a paragraph, numbers, and free speech — then Adapt voice. This learns the shopkeeper’s voice, not product names.

Shop ID

Loading…