whisper-candle

Native OpenAI Whisper in pure Rust. No Python, no PyTorch, no C/C++ bindings.

whisper-candle is a from-scratch port of Whisper's full inference pipeline to Rust on candle. One static binary replaces the reference stack (Python + PyTorch + ffmpeg + numba + tiktoken): audio decoding, log-mel spectrogram, encoder/decoder, greedy & beam-search decoding, word-level timestamps, and subtitle writers — all native.

It is not a wrapper: unlike whisper-rs (whisper.cpp bindings), there is no C or C++ in the dependency tree.

Why

• Zero-dependency deploys — embed speech-to-text in a Rust service or ship a single binary; no Python runtime or conda environment to drag along.
• Verified parity, not "inspired by" — developed test-first against golden fixtures generated from the reference implementation: greedy and beam decoding are token-exact vs PyTorch on the test set, word timestamps within 0.1 s (DESIGN.md documents the test pyramid and tolerances).
• Runs everywhere candle runs — CPU (Accelerate/BLAS), Metal (correct where torch MPS is broken), CUDA behind a feature flag.

Quick start

cargo install whisper-candle-cli
whisper-candle audio.mp3 --model base --output-format srt -o out/

Or from a checkout:

cargo run --release --features accelerate -p whisper-candle-cli -- \
    audio.mp3 --model base --output-format srt -o out/

Models (tiny … large-v3, turbo, .en variants) download automatically from the Hugging Face Hub as safetensors. Useful flags:

--device metal              # Apple GPU
--quantization q4k          # quantize locally to GGUF once (turbo: 0.5 GB vs 1.6 GB)
--word-timestamps           # per-word timing in the json output
--task translate --language ja

As a library:

let device = whisper_core::device("cpu")?;
let mut model = whisper_core::load_model("base", &device)?;
let result = whisper_core::transcribe_file(&mut model, "audio.wav", &Default::default())?;
println!("{}", result.text);

Status

Transcription is at parity with the Python reference (see DESIGN.md for the current phase); remaining work is performance tuning — CPU decoding currently runs ~10–30× real-time on Apple Silicon, within ~4× of PyTorch CPU.

Python appears in this repo only to test the port: tools/ regenerates the golden fixtures and baseline benchmarks from a reference checkout. It is never needed to build or run whisper-candle.

cargo test -p whisper-candle-core                # fast golden tests, no network
cargo test -p whisper-candle-core --release --features accelerate \
    --test model_goldens -- --ignored            # parity suite (downloads whisper-tiny)

Provenance

This port was generated with Claude Code (Claude Fable 5), guided by golden-fixture TDD against openai/whisper.

License

MIT. crates/whisper-core/src/nn.rs derives from candle-transformers (Apache-2.0/MIT); tokenizer vocabularies and mel filterbanks are from openai/whisper (MIT).