Fleet-2
A 30 MB on-device speech-to-text model (Cactus Whistle architecture) for English, German, French, Spanish, Italian, Dutch, Polish and Korean, trained on transcripts written by openai/whisper-large-v3-turbo.
Collection: Speech recognition
Clean text: Fleet-Clean-0.36B chains Fleet-2 with a word tagger that removes fillers, repeats and false starts and fixes punctuation (GGUF).
What changed from Fleet-1
- Teacher: Whisper-large-v3-turbo transcribed new audio, and Fleet-2 learned from those transcripts. This is sequence-level distillation, as in Distil-Whisper. Whisper and Whistle use different tokenizers, so no logits were matched.
- New audio: VoxPopuli (European Parliament speech) for the seven European languages and Zeroth-Korean for Korean. Fleet-1 used FLEURS only.
- Start point: chosen by validation score between Sky-Kim's Korean Whistle and Fleet-1. This run started from
Lucasgitguy-co/Fleet-1.
Evaluation
Each language has 100 short FLEURS test clips. These are the same clips, filtered the same way, as Fleet-1's table. All models ran in the same native Cactus engine with automatic language detection. Lower is better. Whistle has no Korean.
| Language | Metric | Whistle | Sky-Kim | Fleet-1 | Fleet-2 |
|---|---|---|---|---|---|
| en | WER | 13.86% | 14.87% | 14.39% | 14.50% |
| de | WER | 22.45% | 22.71% | 22.18% | 22.02% |
| fr | WER | 22.15% | 25.73% | 25.18% | 24.93% |
| es | WER | 11.83% | 11.83% | 11.93% | 11.93% |
| it | WER | 17.44% | 16.60% | 16.32% | 16.66% |
| nl | WER | 32.95% | 29.39% | 29.53% | 30.06% |
| pl | WER | 45.38% | 43.68% | 43.13% | 41.62% |
| ko | CER | n/a | 33.68% | 52.35% | 42.80% |
Fleet-2 beats Fleet-1 in 4 of 8 languages (de, fr, pl, ko). Korean CER is no worse than Sky-Kim: no. Improvement over Fleet-1 was NOT demonstrated. Check each language; some may regress.
Out-of-domain clips (VoxPopuli / Zeroth-Korean test splits)| Language | Source | Clips | Whistle | Sky-Kim | Fleet-1 | Fleet-2 |
|---|---|---|---|---|---|---|
| en | VoxPopuli | 50 | 10.56% | 12.69% | 10.91% | 11.98% |
| de | VoxPopuli | 50 | 17.32% | 19.66% | 19.53% | 18.36% |
| fr | VoxPopuli | 50 | 13.06% | 12.24% | 12.04% | 13.57% |
| es | VoxPopuli | 50 | 9.99% | 11.89% | 11.70% | 11.04% |
| it | VoxPopuli | 50 | 34.45% | 40.21% | 38.22% | 33.19% |
| nl | VoxPopuli | 50 | 23.06% | 23.29% | 23.17% | 22.95% |
| pl | VoxPopuli | 50 | 22.02% | 23.53% | 22.98% | 23.53% |
| ko | Zeroth-Korean | 50 | n/a | 25.19% | 15.65% | 13.24% |
Fleet-2 trained on other clips from these corpora, so this table shows in-corpus generalisation, not a fully independent test.
These results come from short clips of 12 s or less. They are not full benchmark scores and not significance tests. Accuracy on long, noisy or conversational audio was not measured. Upstream models may have seen FLEURS training data before.
Training
Data: 4326 clips kept out of 6400 that the teacher labeled: 3014 new VoxPopuli/Zeroth clips and 1312 of Fleet-1's FLEURS training clips. All clips are 12 s or less.
Teacher decoding: language forced, temperature 0, no conditioning on previous text. Runtime: the MLX conversion
mlx-community/whisper-large-v3-turbo.Filter: a clip was kept only if all of these held:
- compression ratio ≤ 2.4
- average log-probability ≥ -1.0
- no-speech probability ≤ 0.6
- teacher text within 10% WER of the human transcript (CER for Korean)
The training target is the teacher's text.
Excluded from training: validation and test clips, and any transcript that matches an evaluation transcript.
Optimisation: the start adapter (LoRA plus Korean vocabulary rows) continued with AdamW at 2e-05, warmup/cosine schedule, batch 1, 2 epochs (8652 steps), clip 1.0, W4/A8 quantisation-aware training on CPU (Apple M4).
Checkpoint selection: step 2163, chosen by macro validation error on Fleet-1's FLEURS validation clips (0.2212, against 0.2283 for the start model). Test clips were not used for selection.
Release: published whatever the scores. The comparison above is reported as measured, including any regressions.
training_manifest.json records every clip, teacher transcript and pinned revision. evaluation.json has every score. fleet2.py, teacher.py and run.sh reproduce the run.
Usage
import cactus
from huggingface_hub import hf_hub_download
path = hf_hub_download("Lucasgitguy-co/Fleet-2", "whistle.cact")
print(cactus.transcribe("clip.wav", weights=path)["text"])
Use 16 kHz mono WAV and cactus-compute at revision ee2ffc67fe0f9537d4928e158440c7d228a63100. This is a Cactus model, not a Transformers checkpoint.
License and attribution
Model: Apache License 2.0. The original Whistle LICENSE is kept verbatim. Architecture and pretrained weights: Cactus Compute's Whistle (Mroz et al., 2026). Korean vocabulary and adapter lineage: Sky-Kim/whistle-finetuned-ko and Lucasgitguy-co/Fleet-1. See
UPSTREAM_MODEL_CARD.md.Teacher: openai/whisper-large-v3-turbo, MIT License. Only its transcripts were used; none of its weights are included.
Data:
- VoxPopuli: CC0
- Zeroth-Korean: CC BY 4.0
- FLEURS (Conneau et al., 2022): CC BY 4.0
No audio is included in this repository. See
NOTICE.