parakeet: current int8 v2 model is English-only — document language choice / consider v3 #2

New Issue

2026-04-24T09:56:07-07:00

vh commented

2026-04-24 09:56:07 -07:00

The parakeet stack on irv-ml1 currently pins to sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8, which is English-only. Non-English audio will either fail or produce garbage output.

Current state

stacks/parakeet/.env.example defaults PARAKEET_MODEL_URL to the v2 English-only tarball.
The v3 tarball (sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8) covers 25 European languages at similar size (~400 MB int8) and can be swapped in by bumping PARAKEET_MODEL_URL + wiping *.onnx + tokens.txt from /worktank/parakeet/models/ + restarting.
This is documented in the README under "Switching to the v3 (multilingual) model" but not in .env.example.

What to decide

Keep v2 default for English-centric use. Current state. Lower WER for English than v3 (per NeMo benchmarks) and smaller tokenizer. No change needed beyond making the limitation more prominent in .env.example.
Switch default to v3. Slight English WER regression (typically single-digit relative), but drop-in support for de/es/fr/it/pt/ru and 19 other European languages.
Expose both as parallel stacks. Two containers (parakeet-en, parakeet-multi), two ports, same code. Overkill unless you actually need both live simultaneously.

Next step

Pick an option. If (1), just mention in .env.example that swapping URL requires wiping the models dir — currently only in README. If (2), update the default URL + tag the image rebuild. If (3), design the port allocation + document.

Context

Endpoint: http://10.100.79.3:8765/transcribe
Stack source: stacks/parakeet/ in this repo
Model tarballs: https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models

The parakeet stack on irv-ml1 currently pins to `sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8`, which is **English-only**. Non-English audio will either fail or produce garbage output. ### Current state - `stacks/parakeet/.env.example` defaults `PARAKEET_MODEL_URL` to the v2 English-only tarball. - The v3 tarball (`sherpa-onnx-nemo-parakeet-tdt-0.6b-v3-int8`) covers 25 European languages at similar size (~400 MB int8) and can be swapped in by bumping `PARAKEET_MODEL_URL` + wiping `*.onnx` + `tokens.txt` from `/worktank/parakeet/models/` + restarting. - This is documented in the README under "Switching to the v3 (multilingual) model" but not in `.env.example`. ### What to decide 1. **Keep v2 default for English-centric use.** Current state. Lower WER for English than v3 (per NeMo benchmarks) and smaller tokenizer. No change needed beyond making the limitation more prominent in `.env.example`. 2. **Switch default to v3.** Slight English WER regression (typically single-digit relative), but drop-in support for de/es/fr/it/pt/ru and 19 other European languages. 3. **Expose both as parallel stacks.** Two containers (`parakeet-en`, `parakeet-multi`), two ports, same code. Overkill unless you actually need both live simultaneously. ### Next step Pick an option. If (1), just mention in `.env.example` that swapping URL requires wiping the models dir — currently only in README. If (2), update the default URL + tag the image rebuild. If (3), design the port allocation + document. ### Context - Endpoint: `http://10.100.79.3:8765/transcribe` - Stack source: `stacks/parakeet/` in this repo - Model tarballs: https://github.com/k2-fsa/sherpa-onnx/releases/tag/asr-models

vh added the question label 2026-04-24 09:57:41 -07:00

vh referenced this issue from a commit

2026-08-28 05:52:26 -07:00

memory: run 3c had TWO launches — the third was an untimestamped report

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: vh/esh-pfi-infrastructure#2