Dev build

Latest build of main at a3b5c3b (2026-10-23 23:55 UTC).

v0.3.1

Features

• Added Model::device(), which tells which API (Metal, Vulkan, DX12) and which GPU a loaded model runs on; wavo bench prints it (#49)
• Added a DirectX 12 fallback on Windows machines without a Vulkan driver, such as virtual machines; it shipped in 0.3.0 but was missing from its notes (#46)

Fixes

• Fixed a crash at the first transcription on software renderers such as WARP in Windows virtual machines (#48)

Full Changelog: v0.3.0...v0.3.1

v0.3.0

Features

• Added a fast path for Apple M5 and newer GPUs that runs matrix math on their neural accelerators through the Metal 4 tensor API, about twice as fast as before on GigaAM, Parakeet and Whisper (#43)
• wavo rm and Hub::remove now also delete models that hf keeps in its shared blob store, once no other model uses the file (#41)

Improvements

• Improved Whisper speed on NVIDIA GPUs about 3× and Moonshine's speed and memory everywhere with flash attention (#36)
• Improved Whisper decoding speed on NVIDIA GPUs (#38)
• Improved the first transcription after loading a model on discrete GPUs, e.g. Parakeet on an RTX 4060 from ~100 to ~77 ms (#44)
• Improved speed on integrated and other GPUs without fast kernels: Parakeet about 1.6–2× faster on an Intel N100 (#45)

Docs

• Updated the readme with per-machine speed numbers (Apple M2 and M5, RTX 4060, Intel N100), how wavo is made and tested, and credits (#42)

Full Changelog: v0.2.0...v0.3.0

v0.2.0

Breaking Changes

• Marked the library's Transcript, Token and Error as non-exhaustive, so code outside wavo can no longer build them with struct literals or match them without a wildcard (#21)

Features

• Added wavo AUDIO as a shortcut for wavo run AUDIO; wavo run now offers to download a missing model, and -y downloads it without asking (#29)
• Added Moonshine tiny and base models for fast English transcription (#30)
• Added the other Whisper models: tiny, base, small, medium and their English-only versions, large, large-v2, large-v3 and Breeze-ASR-25 for Taiwanese Mandarin (#20)
• Added model downloads to the library: wavo::hub behind the hub feature finds, downloads and removes models in the Hugging Face cache (#21)
• Added downloading several models at once with wavo pull (#17)
• Added progress bars for wavo pull and wavo run (#18)
• Added prebuilt binaries for macOS, Linux and Windows on GitHub releases, and builds of the latest code in the dev release (#9, #32)

Improvements

• Improved speed on NVIDIA GPUs by 10–40% with tensor cores (#28)
• Improved model load time by about half, so dictation with a model loaded per recording finishes sooner (#24)
• Improved downloads to retry and resume after a dropped connection (#19)

Full Changelog: v0.1.0...v0.2.0

v0.1.0

Features

• Added local speech-to-text with Parakeet TDT v2 and v3, plus GigaAM v3 E2E-RNNT, E2E-CTC, RNNT and CTC models.
• Added Whisper large-v3-turbo for multilingual transcription with segment timestamps (#8).
• Added GPU acceleration through Metal on Apple Silicon and Vulkan on Linux and Windows (#5, #7).
• Added a pure Rust library for loading models and transcribing audio into text and timestamped tokens, with optional CLI dependencies (#2).
• Added pull, list and rm commands to manage models in the shared Hugging Face cache, plus run for offline transcription and bench to measure speed (#2).
• Added WAV, MP3, M4A/AAC, FLAC and Ogg/Vorbis input with automatic downmixing and resampling (#2).
• Added plain text, timestamped JSON and SRT subtitle output, with one subtitle cue per Whisper segment (#2, #10).
• Added automatic splitting of long recordings at pauses, with --segment to control or disable splitting (#3).

Full Changelog: https://github.com/vladkens/wavo/commits/v0.1.0