# Whisper large-v3-turbo OpenAI's [`openai/whisper-large-v3-turbo`](https://huggingface.co/openai/whisper-large-v3-turbo) ported to transcribe.cpp. A 809M-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 (100 languages covered) and emits a transcript in that language; passing `language=""` 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. v3 family adds Cantonese (yue) on top of v2's 99 languages and switches to a 128-bin mel input. See the [upstream model card](https://huggingface.co/openai/whisper-large-v3-turbo) for training data, intended use, and the original evaluation methodology. Licensed Apache-2.0. Ported from upstream commit [`41f01f3`](https://huggingface.co/openai/whisper-large-v3-turbo/commit/41f01f3), pinned 2026-04-25. Validated against the transformers reference at transcribe.cpp commit [`5.6.1`](https://github.com/handy-computer/transcribe.cpp/tree/5.6.1) on 2026-04-26. ## Download | Quantization | Download | Size | WER (LibriSpeech test-clean) | | --- | --- | ---: | ---: | | F16 | [whisper-large-v3-turbo-F16.gguf](https://huggingface.co/handy-computer/whisper-large-v3-turbo-gguf/resolve/main/whisper-large-v3-turbo-F16.gguf) | 1.51 GB | 2.01% | | Q8_0 | [whisper-large-v3-turbo-Q8_0.gguf](https://huggingface.co/handy-computer/whisper-large-v3-turbo-gguf/resolve/main/whisper-large-v3-turbo-Q8_0.gguf) | 845 MB | 2.01% | | Q6_K | [whisper-large-v3-turbo-Q6_K.gguf](https://huggingface.co/handy-computer/whisper-large-v3-turbo-gguf/resolve/main/whisper-large-v3-turbo-Q6_K.gguf) | 660 MB | 2.01% | | Q5_K_M | [whisper-large-v3-turbo-Q5_K_M.gguf](https://huggingface.co/handy-computer/whisper-large-v3-turbo-gguf/resolve/main/whisper-large-v3-turbo-Q5_K_M.gguf) | 591 MB | 2.03% | | Q4_K_M | [whisper-large-v3-turbo-Q4_K_M.gguf](https://huggingface.co/handy-computer/whisper-large-v3-turbo-gguf/resolve/main/whisper-large-v3-turbo-Q4_K_M.gguf) | 511 MB | 2.04% | 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](whisper.md#all-variants). 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](../tools/wer.md) for the harness. ## Quick Start ```bash cmake -B build cmake --build build build/bin/transcribe-cli \ -m models/whisper-large-v3-turbo/whisper-large-v3-turbo-Q8_0.gguf \ samples/jfk.wav ``` If your audio is not already 16 kHz mono WAV, convert it first: ```bash 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) | 286.1 ms (38.4×) | 288.7 ms (38.1×) | | Metal | dots (35.3s) | 649.5 ms (54.4×) | 641.1 ms (55.1×) | | CPU | jfk (11.0s) | 7.60 s (1.4×) | 5.89 s (1.9×) | | CPU | dots (35.3s) | 15.34 s (2.3×) | 11.87 s (3.0×) | macOS 26.4.1, transcribe.cpp `e0fa0f6`. Benchmark reproduction: ```bash uv run scripts/bench/run.py \ --models whisper-large-v3-turbo \ --quants q8_0,q4_k_m \ --samples jfk,dots \ --backends metal,cpu \ --iters 3 --warmup 1 \ --name whisper-large-v3-turbo-publication ``` ### AMD Ryzen 7 PRO 4750U | Backend | Sample | Q8_0 | Q4_K_M | | ------- | ------------ | -------------: | ------------: | | Vulkan | jfk (11.0s) | 4.14 s (2.7×) | 4.15 s (2.7×) | | Vulkan | dots (35.3s) | 8.70 s (4.1×) | 8.88 s (4.0×) | | CPU | jfk (11.0s) | 19.85 s (0.6×) | 15.74 s (0.7×) | | CPU | dots (35.3s) | 40.18 s (0.9×) | 32.22 s (1.1×) | Fedora 43, transcribe.cpp `2ab01b8`. Vulkan device: `AMD Radeon Graphics (RADV RENOIR)`. Benchmark reproduction: ```bash uv run scripts/bench/run.py \ --models whisper-large-v3-turbo \ --quants q8_0,q4_k_m \ --samples jfk,dots \ --backends cpu,vulkan \ --iters 3 --warmup 1 \ --name whisper-large-v3-turbo-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 22 checkpointed tensors fall within per-variant tolerance, and the transcript matches the HF reference verbatim. Tolerance budget lives at [`tests/tolerances/whisper-large-v3-turbo.json`](https://github.com/handy-computer/transcribe.cpp/blob/main/tests/tolerances/whisper-large-v3-turbo.json). Last validated at commit [`1854f57`](https://github.com/handy-computer/transcribe.cpp/tree/1854f57). | Field | Value | | --- | --- | | Reference | transformers 5.6.1 (`WhisperForConditionalGeneration`, CPU fp32) | | Manifest | `tests/golden/whisper/whisper-large-v3-turbo.manifest.json` | | Tolerance file | `tests/tolerances/whisper-large-v3-turbo.json` | | Command | `uv run scripts/validate.py all --family whisper --variant whisper-large-v3-turbo` | 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` | `4.055e-08` | fp32 mixed-radix FFT vs torch fp64 frontend | | `enc.conv1.out` | `1.258e-03` | `4.264e-05` | fp32 conv stem | | `enc.conv2.out` | `2.696e-03` | `5.033e-05` | stride-2 conv stem (matches enc.embed.out) | | `enc.block.0.out` | `1.439e-02` | `1.334e-03` | first encoder block | | `enc.block.31.out` | `2.335e+00` | `3.895e-03` | final encoder block (peak signal grows with depth) | | `enc.final` | `1.343e+00` | `7.386e-03` | 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` | `8.217e-03` | `1.484e-04` | first decoder block, prompt pass | | `dec.block.3.out` | `1.453e-02` | `9.290e-04` | final decoder block (accumulated) | | `dec.out_before_head` | `2.122e-01` | `2.041e-02` | post final LN, pre-vocab projection | | `dec.logits_raw` | `1.462e-01` | `2.968e-02` | vocab projection (raw logits) | | `dec.logits` | `1.771e-01` | `3.328e-02` | log-softmax over vocab | | `dec.logits_raw.gen20` | `3.806e-02` | `6.097e-03` | 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. ```bash uv run --project scripts/envs/whisper \ scripts/convert-whisper.py openai/whisper-large-v3-turbo \ --revision 41f01f3 ``` ### Quantize Run `transcribe-quantize` once per target quant. Example for Q8_0; repeat for the other shipped presets: ```bash build/bin/transcribe-quantize \ models/whisper-large-v3-turbo/whisper-large-v3-turbo-F16.gguf \ models/whisper-large-v3-turbo/whisper-large-v3-turbo-Q8_0.gguf \ --quant Q8_0 ``` ### Validate ```bash uv run scripts/validate.py all --family whisper --variant whisper-large-v3-turbo ``` ### Run real-model tests ```bash cmake -B build -DTRANSCRIBE_BUILD_REAL_MODEL_TESTS=ON cmake --build build TRANSCRIBE_WHISPER_GGUF=$PWD/models/whisper-large-v3-turbo/whisper-large-v3-turbo-Q8_0.gguf \ ctest --test-dir build --output-on-failure -R whisper ```