Models
Hebrish has three fine-tunes. Parakeet is the main model and what the apps use. Nemotron, trained on the same recordings with the same recipe, is experimental and can stream: text appears while you speak. Whisper Small is a phone-sized model for people who already use FUTO Keyboard on Android. Developer notes and suggested implementations for each are below.
One speaker's voice. All three models are trained on recordings of one person: an English speaker in Jerusalem with an Irish accent. How well they work for other voices has not been tested. Help make them multi-speaker.
How v1.0 was trained, exported and benchmarked, with every result and its limits: Hebrish v1.0 white paper (PDF, 9 pages).
| Parakeet fine-tune main | Nemotron fine-tune experimental | Whisper Small fine-tune | |
|---|---|---|---|
| Base model | nvidia/parakeet-tdt-0.6b-v2 | nvidia/nemotron-speech-streaming-en-0.6b | openai/whisper-small.en |
| Architecture | FastConformer encoder, TDT decoder, 0.6B | Cache-aware FastConformer encoder, RNNT decoder, 0.6B | Whisper encoder-decoder, 244M |
| How it transcribes | The whole clip, after you stop speaking | Streaming, in chunks of 80 to 1,120 ms while you speak (also offline) | The whole clip, up to 30 seconds. ACFT lets it process only as much audio as the clip is long |
| Current checkpoint | v1.0 · Jachnun: checkpoint 6, run r6-en30, 2026-10-19 | Checkpoint 2, run r6-nemotron-en30, 2026-10-19 | Run small-en-en30-acft, 2026-10-17 |
| Trained on | Read-Aloud-Voice-Sentences 550bfcaa, 2026-10-19 latest data | Read-Aloud-Voice-Sentences 550bfcaa, 2026-10-19 latest data | Read-Aloud-Voice-Sentences 36845b0d, 2026-10-17 behind |
| Training data | The same for both: 339.2 min of code-switched speech (3,224 clips), 95.3 min of the speaker's own English, 48.5 min of LibriSpeech replay. 20 epochs on one L40S | Earlier and smaller: Read-Aloud-Voice-Sentences as of 2026-10-17, 1,929 clips (215 min: 152.5 code-switching, 40.5 own English, 22.0 public English). 10 epochs | |
| Validation WER (training) | 5.0% | 6.3% (offline decoding) | 3.3% (its own validation set) |
| Benchmarked | Yes: v1.0 73.3% term recall, 3.3% WER on the public benchmark | Public benchmark (offline): 62.2% term recall, 4.7% WER. Streaming at 560 ms: 5.65% CS WER, same as offline | Yes, on its own test set: 5.8% code-switching WER in FUTO's runtime. Not comparable with Parakeet's |
| Formats | .nemo, GGUF, ONNX, sherpa-onnx int8, MLX (still PoC-3) | .nemo only (2.47 GB) | whisper.cpp GGML Q8_0 (264 MB), Transformers fp32 |
| Used by | The Hebrish apps, the browser demo | No app yet | FUTO Keyboard and FUTO Voice Input (Android), imported by hand |
| Licence | CC-BY-4.0 | NVIDIA Open Model License Agreement | MIT |
The Parakeet and Nemotron validation numbers are raw WER on the same 218-clip validation set, measured during training. Whisper's is on its own, smaller validation set. None of them are benchmark results, and the set is larger than earlier versions' sets, so they do not compare with older checkpoints. Benchmarked results: model card.
The more accurate of the two, and the only one with on-device formats. Use it where text can arrive a moment after you stop speaking:
Streaming means the words appear as they are said, with no wait at the end. That opens uses where delay matters more than the last bit of accuracy:
The smallest of the three, at 264 MB. Built for FUTO Keyboard on Android, which runs Whisper models:
Full model card, results and version history: /model. Every format below is v1.0 except MLX.
| Where | Runtime | Files |
|---|---|---|
| Android, Linux, Windows, iOS | sherpa-onnx, offline NeMo transducer, greedy search. What the Hebrish apps use; see Hebrish-Dictation for working Kotlin and Python code | sherpa-onnx int8, ≈ 661 MB |
| Browser | parakeet.js on ONNX Runtime Web (WASM). How /demo works; about 2.3 GB of memory once loaded | ONNX int8 |
| Python | onnx-asr or ONNX Runtime directly | ONNX fp32 and int8 |
| Desktop and Android apps built on transcribe.cpp (Offline Voice Input, Handy) | transcribe.cpp v0.1.3 | GGUF Q8_0 (730 MB), Q4_K_M (475 MB) for phones with 4 GB of RAM or less |
| Mac (Apple Silicon) | pip install parakeet-mlx | MLX still PoC-3 |
| GPU, further fine-tuning | NVIDIA NeMo | .nemo, 2.47 GB |
Audio in is 16 kHz mono. Text out is English, with Hebrew words romanized as in the lexicon, never in Hebrew script.
experimental streaming measured at 560 ms not in any app
Today it runs in NVIDIA NeMo, from the .nemo file. Offline transcription, from the model card:
import nemo.collections.asr as nemo_asr
from huggingface_hub import hf_hub_download
path = hf_hub_download(
"Hebrish/Nemotron-Speech-Streaming-EN-0.6B-English-Hebrew-Codeswitching",
"Nemotron-Speech-Streaming-EN-0.6B-English-Hebrew-Codeswitching.nemo",
)
model = nemo_asr.models.ASRModel.restore_from(path)
print(model.transcribe(["clip.wav"])[0].text) # 16 kHz monoStreaming works as for the base model, with cache-aware inference. Set att_context_size = [70, N]:
| N | 0 | 1 | 6 | 13 |
|---|---|---|---|---|
| Chunk | 80 ms | 160 ms | 560 ms | 1,120 ms |
Smaller chunks mean less delay. NeMo's speech_to_text_cache_aware_streaming_infer.py example script shows the loop; see the base model card.
Nemotron fine-tune on Hugging Face
A full fine-tune of openai/whisper-small.en, then FUTO's ACFT procedure. FUTO Keyboard runs Whisper with a dynamic audio context: it only processes as much of the 30-second window as the clip is long, which makes short dictations fast. Stock and ordinary fine-tuned Whisper models break under that setting, repeating themselves or never stopping. ACFT fixes it. The base is English-only, so it never writes Hebrew script.
Training data. Read-Aloud-Voice-Sentences at revision 36845b0d (2026-10-17), the same parent dataset the Parakeet and Nemotron fine-tunes use, but two days earlier: 152.5 minutes of code-switched speech against 339.2 for their latest checkpoints. Whisper Hebrish goes back further: its first version (2025-11-17) was trained on 516 desktop clips in English-Hebrew-Mixed-Sentences, recorded that day. 498 of those clips were later brought into Read-Aloud-Voice-Sentences with their spellings made canonical, so they are part of what every current model trains on.
FUTO Keyboard (Android): download ggml-whisper-hebrish-small-en-acft-q8_0.bin (264 MB) on the phone. Open it with FUTO Keyboard, or go to Settings → Languages & Models → Import resource file, and choose English when asked. Revert to Default on the same screen undoes it. These steps come from FUTO Keyboard's source code and were not checked on a phone.
whisper.cpp, with the same dynamic context FUTO uses (for a 6-second clip, 6 × 50 + 32 = 332):
whisper-cli -m ggml-whisper-hebrish-small-en-acft-q8_0.bin -f clip.wav -bs 5 -ac 332Transformers:
from transformers import pipeline
asr = pipeline("automatic-speech-recognition", model="Hebrish/Whisper-Hebrish-Small-ACFT")
print(asr("clip.wav")["text"]) # English-only model: do not pass language/taskResults from the model card, in FUTO's own runtime (q8_0, beam 5), on a 155-sentence code-switching test set whose Hebrew terms were never trained on:
| Model | Code-switching WER | Unseen-term recall | Own English WER |
|---|---|---|---|
| FUTO's stock English-244 model | 11.8% | 26% | 4.1% |
| Whisper Hebrish Small ACFT | 5.8% | 50% | 2.8% |
This test set is not the one Parakeet's benchmark uses, so these numbers do not compare with Parakeet's. Clips longer than 30 seconds score much worse (21 to 27% WER). A larger sibling, Whisper Hebrish on whisper-large-v3-turbo (809M), scored 5.2% on the same set at full context. Both are MIT-licensed.
Whisper Small fine-tune on Hugging Face