Skip to content

whisper-small.md

Latest commit

 

History

History
180 lines (137 loc) · 8.03 KB

whisper-small.md

File metadata and controls

180 lines (137 loc) · 8.03 KB
Raw

Whisper small

OpenAI's openai/whisper-small ported to transcribe.cpp. A 244M-parameter encoder-decoder transformer (audio encoder + autoregressive text decoder with cross-attention).

What it's for

Offline multilingual speech-to-text and any-language → English speech translation. The model auto-detects the audio's language (99 languages covered) and emits a transcript in that language; passing language="<code>" and task="translate" to the underlying whisper_full_params produces an English translation instead. transcribe-cli reads a 16 kHz mono WAV and returns the transcript text. Long audio is handled via 30-second chunked decoding.

See the upstream model card for training data, intended use, and the original evaluation methodology.

Licensed Apache-2.0. Ported from upstream commit 973afd2, pinned 2026-04-25. Validated against the transformers reference at transcribe.cpp commit 5.6.1 on 2026-04-26.

Download

Quantization Download Size WER (LibriSpeech test-clean)
F32 whisper-small-F32.gguf 924 MB 3.33%
F16 whisper-small-F16.gguf 470 MB 3.34%
Q8_0 whisper-small-Q8_0.gguf 257 MB 3.33%
Q6_K whisper-small-Q6_K.gguf 202 MB 3.33%
Q5_K_M whisper-small-Q5_K_M.gguf 185 MB 3.37%
Q4_K_M whisper-small-Q4_K_M.gguf 164 MB 3.40%

WER measured on the full LibriSpeech test-clean split (2620 utterances) with transcribe.cpp's default greedy decode and segment timestamps enabled — the same runs summarized in the Whisper family table. Numbers come from a single Metal-backed run; Metal's non-deterministic parallel reductions add ~0.1pp of run-to-run variance on the noise floor, and quantization is otherwise generally WER-neutral. See the WER methodology for the harness.

Quick Start

cmake -B build
cmake --build build

build/bin/transcribe-cli \
  -m models/whisper-small/whisper-small-Q8_0.gguf \
  samples/jfk.wav

If your audio is not already 16 kHz mono WAV, convert it first:

ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav

Performance

Cells are wall-clock latency (mel + encode + decode, mean over the recorded iterations after warmup), with speedup over realtime in parentheses. Units: ms below 1 s, s above (2 decimal places). Decode latency dominates as model size grows; the encoder is only run once per 30-second window.

Apple M4 Max

Backend Sample Q8_0 Q4_K_M
Metal jfk (11.0s) 113.1 ms (97.2×) 113.5 ms (96.9×)
Metal dots (35.3s) 349.3 ms (101.2×) 340.0 ms (103.9×)
CPU jfk (11.0s) 1.43 s (7.7×) 1.30 s (8.4×)
CPU dots (35.3s) 3.01 s (11.8×) 2.74 s (12.9×)

macOS 26.4.1, transcribe.cpp e0fa0f6.

Benchmark reproduction:

uv run scripts/bench/run.py \
  --models whisper-small \
  --quants q8_0,q4_k_m \
  --samples jfk,dots \
  --backends metal,cpu \
  --iters 3 --warmup 1 \
  --name whisper-small-publication

AMD Ryzen 7 PRO 4750U

Backend Sample Q8_0 Q4_K_M
Vulkan jfk (11.0s) 1.03 s (10.6×) 0.96 s (11.4×)
Vulkan dots (35.3s) 2.57 s (13.7×) 2.47 s (14.3×)
CPU jfk (11.0s) 3.95 s (2.8×) 3.27 s (3.4×)
CPU dots (35.3s) 8.91 s (4.0×) 7.47 s (4.7×)

Fedora 43, transcribe.cpp e0fa0f6. Vulkan device: AMD Radeon Graphics (RADV RENOIR).

Benchmark reproduction:

uv run scripts/bench/run.py \
  --models whisper-small \
  --quants q8_0,q4_k_m \
  --samples jfk,dots \
  --backends cpu,vulkan \
  --iters 3 --warmup 1 \
  --name whisper-small-publication

Numerical Validation

transcribe.cpp is validated tensor-by-tensor against the transformers reference (WhisperForConditionalGeneration, fp32 CPU) on the manifest's cases (samples/jfk.wav and samples/german.wav). All 23 checkpointed tensors fall within per-variant tolerance, and the transcripts match the HF reference verbatim. Tolerance budget lives at tests/tolerances/whisper-small.json. Last validated at commit 1854f57.

Field Value
Reference transformers 5.6.1 (WhisperForConditionalGeneration, CPU fp32)
Manifest tests/golden/whisper/whisper-small.manifest.json
Tolerance file tests/tolerances/whisper-small.json
Command uv run scripts/validate.py all --family whisper --variant whisper-small

Selected tensors (worst observed across cases; see tolerance file for per-tensor budgets):

Tensor Max abs diff Mean abs diff Notes
enc.mel.in 2.229e-05 3.381e-08 fp32 mixed-radix FFT vs torch fp64 frontend
enc.conv1.out 5.454e-06 4.815e-08 fp32 conv stem
enc.conv2.out 2.146e-05 2.704e-07 stride-2 conv stem (matches enc.embed.out)
enc.block.0.out 2.134e-05 6.858e-07 first encoder block
enc.block.11.out 1.669e-02 6.139e-06 final encoder block (peak signal grows with depth)
enc.final 2.100e-03 3.882e-06 post-LN encoder output
dec.token_emb 0.000e+00 0.000e+00 exact zero-drift (ggml_get_rows on the F32 GGUF)
dec.block.0.out 4.768e-06 2.997e-07 first decoder block, prompt pass
dec.block.11.out 3.662e-04 2.471e-06 final decoder block (accumulated)
dec.out_before_head 1.755e-04 7.318e-06 post final LN, pre-vocab projection
dec.logits_raw 3.719e-05 7.083e-06 vocab projection (raw logits)
dec.logits 8.965e-05 1.926e-05 log-softmax over vocab
dec.logits_raw.gen20 2.289e-05 8.815e-06 step-20 logits (KV-cached path)

The C++ mel frontend (Slaney filterbank + Hann periodic window + whisper-style log-mel compression) drives enc.mel.in to fp32-vs-fp64 STFT precision drift; downstream tensors stay within budget. KV-cached decoder runs through F16 self/cross caches by default — flip with --kv-type f32 for tighter parity.

Reproduction

Convert

The whisper converter loads from a Hugging Face checkpoint and emits a reference-dtype GGUF.

uv run --project scripts/envs/whisper \
  scripts/convert-whisper.py openai/whisper-small \
  --revision 973afd2

Quantize

Run transcribe-quantize once per target quant. Example for Q8_0; repeat for the other shipped presets:

build/bin/transcribe-quantize \
  models/whisper-small/whisper-small-F32.gguf \
  models/whisper-small/whisper-small-Q8_0.gguf \
  --quant Q8_0

Validate

uv run scripts/validate.py all --family whisper --variant whisper-small

Run real-model tests

cmake -B build -DTRANSCRIBE_BUILD_REAL_MODEL_TESTS=ON
cmake --build build

TRANSCRIBE_WHISPER_GGUF=$PWD/models/whisper-small/whisper-small-Q8_0.gguf \
  ctest --test-dir build --output-on-failure -R whisper