Jevvellabs/whisper-large-v3-turbo-cs-de-en-es-pl)A fine-tuned variant of openai/whisper-large-v3-turbo optimized by Jevvellabs AI Services for German, English, Spanish, and Polish intra-sentence and inter-sentence code-switching, while preserving near-baseline monolingual human ASR accuracy and zero silence hallucinations.
Both Hugging Face Transformers (model.safetensors, FP16) and a ready-to-run whisper.cpp GGML Q8_0 (ggml/ggml-whisper-large-v3-turbo-iter2-q8_0.bin, 834 MiB) artifact are included in this repository.
whisper.cpp v1.9.4)All evaluations were run directly on the final merged & Q8_0-quantized GGML artifact inside whisper.cpp (-l auto, Silero VAD v5.1.2) against stock ggml-large-v3-turbo-q8_0.bin using paired bootstrap 95% confidence intervals:
| Benchmark | Stock whisper-large-v3-turbo (Q8_0) |
This Model (iter2-q8_0) |
Delta |
|---|---|---|---|
Held-out DE/EN/ES/PL Code-Switching WER (n=24, voice-disjoint) |
30.48% | 16.52% | -13.96 pts (-45.8% rel) · 95% CI [-22.07%, -6.80%] |
Technical Entity Recall (n=24, held-out) |
60.42% | 75.00% | +14.58 pts |
General Monolingual google/fleurs WER (n=48: DE/EN/ES/PL) |
6.33% | 6.65% | +0.32 pts · 95% CI [-0.39%, +1.04%] (statistically indistinguishable from baseline) |
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) |
9.01% | 9.46% |
English (en_us) |
8.80% | 8.80% |
Spanish (es_419) |
2.56% | 2.56% |
Polish (pl_pl) |
5.42% | 6.40% |
Full machine-readable gate metrics are included in benchmark_gate_report.json.
whisper-large-v3-turbo uses a 32-layer encoder paired with a thin 4-layer decoder. Keeping the entire 32-layer encoder frozen protects general acoustic generalization across unseen human speakers while focusing LoRA capacity (q_proj, k_proj, v_proj, out_proj, fc1, fc2) on multilingual decoding stability, followed by full weight merge (merge_and_unload()) and Q8_0 quantization via whisper-quantize.<|de|>, <|en|>, <|es|>, <|pl|>) so the thin decoder transcribes foreign-language clauses verbatim instead of forcing phonetic transliteration or translation under -l auto.google/fleurs replay across DE, EN, ES, and PL.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-large-v3-turbo-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 — Q8_0, 834 MiB)Download ggml/ggml-whisper-large-v3-turbo-iter2-q8_0.bin (833.7 MiB, SHA-256 384fdf0b0da6604816a8c3743e4f1ddb1ef0578fd3d470043c0620d9b5a5457f):
huggingface-cli download Jevvellabs/whisper-large-v3-turbo-cs-de-en-es-pl \
ggml/ggml-whisper-large-v3-turbo-iter2-q8_0.bin --local-dir .
./whisper-cli -m ggml/ggml-whisper-large-v3-turbo-iter2-q8_0.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 lower memory footprint and faster CPU inference (465 MiB F16, 18.52% CS WER), see Jevvellabs/whisper-small-cs-de-en-es-pl.
Released under the MIT License, matching upstream openai/whisper-large-v3-turbo.
model.safetensors 1.6 GBggml/ggml-whisper-large-v3-turbo-iter2-q8_0.bin 874 MBopenai/whisper-large-v3-turbo