license: apache-2.0
base_model: Sky-Kim/whistle-finetuned-ko
language:
- en
- de
- fr
- es
- it
- nl
- pl
- ko
pipeline_tag: automatic-speech-recognition
datasets:
- google/fleurs
tags:
- whistle
- cactus
- lora
- speech-to-text
Technical report: Read paper 路 PDF 路 BibTeX 路 All models
Collection: Speech recognition
Fleet-1
Experimental continuation of Sky-Kim/whistle-finetuned-ko, retaining its trained Korean vocabulary, language tags, and LoRA weights. Original architecture and pretrained weights: Cactus Compute's Whistle.
Improvement across all eight languages was NOT demonstrated. Inspect the per-language results; some may regress.
Evaluation
Automatic language detection; same native Cactus engine for both models. Lower is better.
| Language | Metric | Source | Fleet-1 |
|---|---|---|---|
| en | WER | 14.87% | 14.39% |
| de | WER | 22.71% | 22.18% |
| fr | WER | 25.73% | 25.18% |
| es | WER | 11.83% | 11.93% |
| it | WER | 16.60% | 16.32% |
| nl | WER | 29.39% | 29.53% |
| pl | WER | 43.68% | 43.13% |
| ko | CER | 33.68% | 52.35% |
These are 100 short FLEURS test clips per language, filtered to at most 12 seconds and 158 text tokens, not full benchmark scores. They do not establish statistical significance or accuracy on long, noisy, or conversational recordings. Test clips were not used for this run's training or checkpoint selection. The upstream model was previously evaluated on FLEURS and trained on some FLEURS training clips. Prior training exposure in the original Cactus model is not fully known.
Training
400 clips per language from FLEURS train; 50 per language from validation. Original splits are kept separate; repeated normalized transcripts within a language are excluded across selected splits. First eligible unique rows in dataset order; epoch order shuffled with seed 42. Existing adapter is continued with a fresh AdamW optimizer, learning rate 2e-05, batch 1, two epochs, gradient clipping 1, warmup/cosine schedule, W4/A8 quantization-aware training on CPU. Best trained epoch by macro-average validation loss: 2. Training selection is not based on test scores. No guarantee of improvement over the source.
Source revision: b8a4f3188a828a0e87e2713b2d3e83d71d893068. Training code revision: ee2ffc67fe0f9537d4928e158440c7d228a63100. Exact examples and settings: training_manifest.json. Language-detection scores: evaluation.json.
Usage
import cactus
from huggingface_hub import hf_hub_download
path = hf_hub_download("Lucasgitguy-co/Fleet-1", "whistle.cact")
print(cactus.transcribe("clip.wav", weights=path))
Use 16 kHz mono WAV audio and the Cactus revision recorded above. This is a Cactus model, not a Transformers AutoModel checkpoint.
License and attribution
Model: Apache License 2.0; original LICENSE retained verbatim. This derivative modifies the source adapter through further multilingual training. See UPSTREAM_MODEL_CARD.md for the source author's training details, attribution, and citation to Whistle (Mroz et al., 2026, Cactus Compute).
Training/evaluation data: FLEURS, FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech, Conneau et al. (2022), CC BY 4.0. Audio was converted to 16 kHz mono PCM and subsetted. Source-model data also includes Zeroth-Korean, CC BY 4.0. No training audio is included in this model repository.