Jevvellabs/whisper-small-cs-de-en-es-pl)A fine-tuned variant of openai/whisper-small optimized by Jevvellabs AI Services for German, English, Spanish, and Polish intra-sentence and inter-sentence code-switching, while preserving monolingual ASR accuracy and zero silence hallucinations.
Both Hugging Face Transformers (model.safetensors) and a ready-to-run whisper.cpp GGML F16 (ggml/ggml-whisper-small-iter2-f16.bin) artifact are included in this repository.
whisper.cpp v1.9.4)All evaluations were run directly on the final merged GGML artifact inside whisper.cpp (-l auto, Silero VAD v5.1.2) against stock ggml-small.bin using paired bootstrap 95% confidence intervals:
| Benchmark | Stock openai/whisper-small |
This Model (iter2-f16) |
Delta |
|---|---|---|---|
Held-out DE/EN/ES/PL Code-Switching WER (n=24, voice-disjoint) |
31.62% | 18.52% | -13.11 pts (-41.4% rel) · 95% CI [-20.52%, -5.82%] |
Technical Entity Recall (n=24, held-out) |
64.58% | 77.08% | +12.50 pts |
General Monolingual google/fleurs WER (n=48: DE/EN/ES/PL) |
11.71% | 12.13% | +0.42 pts · 95% CI [-1.12%, +2.06%] (no significant regression) |
Silence Hallucination Rate (n=4 silence clips) |
0.0% | 0.0% | 0.0% |
google/fleurs Breakdown (n=12 per language)| Language | Stock Baseline WER | This Model WER |
|---|---|---|
German (de_de) |
14.86% | 18.02% |
English (en_us) |
13.20% | 11.60% |
Spanish (es_419) |
7.69% | 8.42% |
Polish (pl_pl) |
11.82% | 11.33% |
Full machine-readable gate metrics are included in benchmark_gate_report.json.
q_proj, k_proj, v_proj, out_proj, fc1, fc2) and merged into full weights (merge_and_unload()).<|de|>, <|en|>, <|es|>, <|pl|>) so the decoder stops translating or hallucinating foreign clauses when -l auto selects a single dominant language tag for a mixed-language segment.google/fleurs replay across all four languages.google/fleurs (CC-BY-4.0) and synthetic voice-disjoint TTS utterances. Zero private, proprietary, or client audio was used.transformers (Python)import torch
from transformers import pipeline
asr = pipeline(
"automatic-speech-recognition",
model="Jevvellabs/whisper-small-cs-de-en-es-pl",
dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
device=0 if torch.cuda.is_available() else -1,
)
result = asr("mixed_de_en_es_pl_audio.wav")
print(result["text"])
whisper.cpp (CPU / Edge)Download ggml/ggml-whisper-small-iter2-f16.bin (465 MiB, SHA-256 fc6c1bad6368a4abc558e7d7ee24befb1a612277dd244e4631254528ecb60f15):
huggingface-cli download Jevvellabs/whisper-small-cs-de-en-es-pl \
ggml/ggml-whisper-small-iter2-f16.bin --local-dir .
./whisper-cli -m ggml/ggml-whisper-small-iter2-f16.bin -l auto -sns -f audio.wav
Tip for multi-turn multilingual recordings: Combine
-l autowith Silero VAD (--vad) and pause-based segmentation (>= 0.50ssilence split, merging short turns< 2.5s) for optimal mixed-language accuracy and latency.
For even lower WER on German, English, Spanish, and Polish code-switching (16.52% CS WER, 6.65% FLEURS WER), see Jevvellabs/whisper-large-v3-turbo-cs-de-en-es-pl.
Released under the MIT License, matching upstream openai/whisper-small.
ggml/ggml-whisper-small-iter2-f16.bin 488 MBmodel.safetensors 484 MBopenai/whisper-small