Whisper large

OpenAI's openai/whisper-large ported to transcribe.cpp. A 1.55B-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 4ef9b41, 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-large-F32.gguf	5.75 GB	2.72%
F16	whisper-large-F16.gguf	2.89 GB	2.74%
Q8_0	whisper-large-Q8_0.gguf	1.55 GB	2.74%
Q6_K	whisper-large-Q6_K.gguf	1.21 GB	2.62%
Q5_K_M	whisper-large-Q5_K_M.gguf	1.08 GB	2.70%
Q4_K_M	whisper-large-Q4_K_M.gguf	950 MB	2.67%

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-large/whisper-large-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)	476.5 ms (23.1×)	465.1 ms (23.6×)
Metal	dots (35.3s)	1.33 s (26.6×)	1.26 s (28.0×)
CPU	jfk (11.0s)	9.63 s (1.1×)	7.43 s (1.5×)
CPU	dots (35.3s)	19.88 s (1.8×)	15.49 s (2.3×)

macOS 26.4.1, transcribe.cpp e0fa0f6.

Benchmark reproduction:

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

AMD Ryzen 7 PRO 4750U

Backend	Sample	Q8_0	Q4_K_M
Vulkan	jfk (11.0s)	6.26 s (1.8×)	6.13 s (1.8×)
Vulkan	dots (35.3s)	14.41 s (2.5×)	13.72 s (2.6×)
CPU	jfk (11.0s)	26.18 s (0.4×)	19.83 s (0.6×)
CPU	dots (35.3s)	55.64 s (0.6×)	43.98 s (0.8×)

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

Benchmark reproduction:

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

Numerical Validation

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

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

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	2.146e-06	3.695e-08	fp32 conv stem
enc.conv2.out	2.098e-05	3.419e-07	stride-2 conv stem (matches enc.embed.out)
enc.block.0.out	2.718e-05	8.861e-07	first encoder block
enc.block.31.out	3.262e-01	7.382e-06	final encoder block (peak signal grows with depth)
enc.final	1.916e-03	5.219e-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	1.907e-05	3.328e-07	first decoder block, prompt pass
dec.block.31.out	2.441e-04	6.015e-06	final decoder block (accumulated)
dec.out_before_head	2.522e-04	1.035e-05	post final LN, pre-vocab projection
dec.logits_raw	9.346e-05	1.508e-05	vocab projection (raw logits)
dec.logits	8.965e-05	2.757e-05	log-softmax over vocab
dec.logits_raw.gen20	3.052e-05	1.148e-05	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-large \
  --revision 4ef9b41

Quantize

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

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

Validate

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

Run real-model tests

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

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