Falafel / README.md
danielrosehill's picture
Status banner (outdated or current) and auto-approval gate text
8b6abec verified
|
Raw History Blame Contribute Delete
5.54 kB
metadata
license: cc-by-nc-4.0
language:
  - en
base_model: nvidia/parakeet-tdt-0.6b-v2
base_model_relation: finetune
pipeline_tag: automatic-speech-recognition
library_name: nemo
tags:
  - parakeet
  - tdt
  - speaker-adaptation
  - english
  - hebrew
  - code-switching
  - voice-commands
  - command-recognition
  - dictation
datasets:
  - danielrosehill/Read-Aloud-Voice-Sentences
  - openslr/librispeech_asr
extra_gated_prompt: >-
  Access is approved automatically once you accept the terms. Use is governed by
  the Hebrish terms of use (https://hebrish.app/terms): non-commercial use is
  free; commercial use needs a commercial licence from access@hebrish.app.
extra_gated_fields:
  Name: text
  Organisation: text
  Intended use: text
  I agree to use this model under the Hebrish terms of use (non-commercial unless I hold a commercial licence): checkbox

Falafel v1.0: Parakeet TDT 0.6B v2 for spoken commands to a computer or phone

New to Hebrish? The Hebrish Index explains the project and which model to use: Sabich Rolling for dictation, Falafel for voice commands.

Falafel is Hebrish's speech-recognition model for short spoken commands to a computer or phone, much of it English-Hebrew code-switching, in one speaker's voice. This is Falafel v1.0, the desktop model, a fine-tune of Parakeet TDT 0.6B v2.

Name. Falafel is the code name of Hebrish's voice-command line; later retrains keep it and bump the version. This is the desktop model; its small sibling for phone and laptop is Falafel 110M. The training run id is r38-falafel. Device: the desktop (.nemo with NeMo, e.g. on ROCm; the int8 build runs on CPU). CDN manifest: https://cdn.hebrish.app/models/falafel/latest.json.

A full fine-tune of nvidia/parakeet-tdt-0.6b-v2 on one speaker's read-aloud phone and desktop recordings (an Irish English speaker living in Israel), run r38-falafel, trained 2026-11-10 on one L40S (Modal) from stock. Output keeps casing and punctuation; numbers are written in digits.

Other Hebrish models: Falafel 110M v1.0 (phone and laptop) · Sabich, English-Hebrew code-switching, tech and general dictation.

Access and terms of use. Governed by the Hebrish terms of use; access is by request (button on this page, or access@hebrish.app). Non-commercial use is free; commercial use needs a commercial licence.

Results

Not benchmarked. Best validation WER 3.58% (best-epoch=13-val_wer=0.0358.ckpt) on primary_val: 325 held-out command takes, 11.4 min, from one speaker. It picked the checkpoint, so it is an indication of fit, not a benchmark and not comparable with other models' WER.

Training data

Source: danielrosehill/Read-Aloud-Voice-Sentences at revision da82b58d626a; the mix is rebuilt from it by prepare (blend commands).

  • Rows tagged voice-commands (4,816 rows, 170.3 min of the speaker's takes of short spoken commands to a computer or phone, much of it English-Hebrew code-switching; sentences in danielrosehill/Voice-Command-Sentences), plus 28.4 min of his own untagged English and LibriSpeech-PC replay at 2:1 (English share 0.20 of the blend); 21.3 min of that total is priority-word rows repeated 3x. Every other task tag is left out.
  • Mix: 220.1 min of the speaker's takes plus 14.4 min of LibriSpeech-PC replay (cased, punctuated; CC-BY-4.0) = 234.5 min, 5,519 clips.
  • Validation: 1 in 16 of the speaker's train takes by id hash, used only to pick the checkpoint.

Training

NeMo 2.7.3 on one L40S (Modal), bf16. AdamW, lr 3e-05 (cosine, 10% warmup), 20 epochs, batch 16, seed 7. Best checkpoint best-epoch=13-val_wer=0.0358.ckpt, 47.1 min of training.

The speaker's share of 220.1 min is 170.3 min of command takes, 28.4 min of his own untagged English and 21.3 min of priority-word repeats (363 rows, repeated 3x).

Usage

import nemo.collections.asr as nemo_asr
from huggingface_hub import hf_hub_download
path = hf_hub_download("Hebrish/Falafel", "Falafel.nemo")
model = nemo_asr.models.ASRModel.restore_from(path)
print(model.transcribe(["clip.wav"])[0].text)   # 16 kHz mono

Builds: Hebrish/Falafel-sherpa-onnx-int8 (sherpa-onnx int8, offline transducer, what the Hebrish apps download from the CDN manifest https://cdn.hebrish.app/models/falafel/latest.json). No GGUF build.

Limitations

  • One speaker. It is adapted to one voice, accent and phone; others will see less benefit.
  • Tuned for short spoken commands to a computer or phone; long-form dictation and general prose are not what it is for (use Sabich or Sabich-Rolling).
  • Not benchmarked. The only number is the best validation WER (3.58%) on primary_val, 325 held-out command takes (11.4 min) picked by id hash from the same dataset revision. It selected the checkpoint and is not an independent test.

Licence

CC-BY-NC-4.0 (non-commercial), with use governed by the Hebrish terms above. Base model nvidia/parakeet-tdt-0.6b-v2 by NVIDIA, licensed under the CC-BY-4.0. Training data Read-Aloud-Voice-Sentences and LibriSpeech-PC are CC-BY-4.0.