whisper-hu-small-finetuned — GGML (whisper.cpp)

GGML quantizations of sarpba/whisper-hu-small-finetuned, a Hungarian fine-tune of openai/whisper-small (12 encoder + 12 decoder layers, 768 d, 80 mel bins, 51865-token multilingual vocab), packaged for whisper.cpp (ggml-*.bin, the legacy GGML container that whisper-cli / whisper-server load — not GGUF, and not a transformers checkpoint).

Magyarul, röviden: a magyar whisper-small fine-tune whisper.cpp-hez készült GGML kvantált változatai (f16, q8_0, q5_0, q4_0). A q8_0 gyakorlatilag veszteségmentes (+0,005 WER-pont a f16-hoz képest), a q5_0 +0,13, a q4_0 +0,26 pontot veszít, miközben a modell 1,8–3,4× kisebb. Ez a modell a szövegfelismerésre való: a saját időbélyegei ebben a runtime-ban használhatatlanok (minden szegmens 0,00→30,00), szószintű időzítéshez külső aligner kell.

Files

Which one? q8_0 is the recommended default (quality-neutral vs f16, 1.8× smaller). q5_0 is a good size/quality compromise; q4_0 is 3.4× smaller than f16 for +0.26 WER points. f16 is the exact conversion reference.

Measured quality and speed

Protocol: 1 000 utterances of CommonVoice-22 Hungarian test split, 1.589 h of audio, 16 kHz mono, greedy decoding (-bs 1 -bo 1 -nf), whisper-server with concurrency 1, RTX 3090, whisper.cpp v1.9.5 (commit 231638a, ggml 0.27.0, CUDA 13.0). Text normalized with the benchmark's Hungarian normalizer (NORM_V=2: lowercase, punctuation stripped, accents kept — á é í ó ő ú ű are not folded). WER/CER are percentages.

For reference, the same checkpoint in the PyTorch runtime (transformers, fp16, batch 8, greedy) scores 19.04 % WER / 4.25 % CER at 102.7× on the same 1 000 utterances. The GGML f16 hypotheses are 90.9 % word-for-word identical to the PyTorch fp16 hypotheses (>97.9 % within 95 % character similarity; runtime-vs-runtime WER 2.0 %), so the conversion is faithful; the aggregate difference comes from whisper.cpp's segment/no-speech handling, not from the conversion.

Quantization costs almost nothing on this model: q8_0 is free, and even q4_0 costs 0.26 WER points. (Note: this is a 244 M-parameter model — at this size the encoder dominates and the x-factor moves only a few percent between levels, and not monotonically in this run. Treat quantization here as a size/RAM lever, not a speed lever.)

How it compares to the sibling releases

Same pool, same protocol, q8_0 pick:

Usage

git clone https://github.com/ggml-org/whisper.cpp
cd whisper.cpp && cmake -B build -DGGML_CUDA=1 && cmake --build build -j --config Release

hf download sarpba/whisper-hu-small-finetuned-GGML \
  ggml-whisper-hu-small-finetuned-q8_0.bin --local-dir .
ffmpeg -i input.mp3 -ar 16000 -ac 1 -c:a pcm_s16le audio.wav

./build/bin/whisper-cli \
  -m ggml-whisper-hu-small-finetuned-q8_0.bin -l hu -f audio.wav

Useful flags: -bs 1 -bo 1 -nf greedy (what the WER table used; the CLI default is beam search -bs 5 -bo 5); -dev 0 GPU selection; -ng = no GPU (CPU) in this build, -t 8 CPU threads; -mc 0 drops text context (~25 % faster on short utterances); -oj / -ojf / -osrt output formats.

How these files were produced

# source checkpoint, pinned
hf download sarpba/whisper-hu-small-finetuned --revision 695bec9fd9ac32998ade9cfe59e7e486695f7339 --local-dir src

# HF transformers names -> OpenAI/ggml tensor contract (479 tensors, config-driven dims)
python conversion/hf_to_openai_pt.py src model-openai-f16.pt

# .pt -> legacy ggml .bin (needs openai/whisper's mel_filters.npz + multilingual.tiktoken)
python whisper.cpp/models/convert-pt-to-ggml.py model-openai-f16.pt openai-whisper .
whisper.cpp/build/bin/whisper-quantize ggml-model.bin out-q8_0.bin q8_0
whisper.cpp/build/bin/whisper-quantize ggml-model.bin out-q5_0.bin q5_0
whisper.cpp/build/bin/whisper-quantize ggml-model.bin out-q4_0.bin q4_0

The produced GGML tensor list is 479/479 tensors of the whisper.cpp loader contract (11 + 15·12 + 24·12 for an encoder-12/decoder-12 model), with n_mels=80, n_vocab=51865, f16 matmul/conv/token weights and f32 biases/LayerNorms/positional embeddings — the same dtype policy the prebuilt ggerganov bins use. Tensor-count equality and the hypothesis-agreement check above are the conversion's correctness proof.

Limitations

• No reliable timestamps from this model in whisper.cpp. Every segment comes out pinned to 0.00 → 30.00 regardless of audio length (verified on 5 s and 10.5 s clips; the same happens with the owner's turbo fine-tune and the base-soup model, while base OpenAI Whisper has sane segments in the same build). Use this model for text, and get word timings from a forced aligner (e.g. a Hungarian wav2vec2 CTC model or MFA).
• Hungarian read-speech domain. Numbers are on CommonVoice-22 HU (clean read speech). Expect higher error on noisy, telephone or conversational audio. Punctuation and rare proper nouns are not guaranteed.
• Not loadable by transformers, faster-whisper or CTranslate2. Use the original repo for PyTorch weights.
• This is the legacy GGML .bin container, not GGUF. whisper.cpp does not consume GGUF for Whisper models; do not rename these files to .gguf.

License

The source model card declares Apache-2.0; this quantized derivative is released under the same terms. The upstream base model (openai/whisper-small) is MIT. See LICENSE.

Provenance

• Source: sarpba/whisper-hu-small-finetuned @ 695bec9fd9ac32998ade9cfe59e7e486695f7339
• Base: openai/whisper-small
• Toolchain: whisper.cpp v1.9.5 (231638a0717b7822784297e5980965ffcec05810), ggml 0.27.0
• Arch: 244 M params, 80 mel, 1500 audio ctx, 448 text ctx, 51865 vocab
• Produced on 2026-10-25; rows in benchmark.json.