Skip to content

Latest commit

 

History

History
80 lines (64 loc) · 3.92 KB

whisper-cpp.md

File metadata and controls

80 lines (64 loc) · 3.92 KB

VoiceStudio — whisper.cpp Speech Recognition

Optional local transcription and translation through the upstream whisper.cpp CLI. It runs on the CPU everywhere and on the GPU when your CLI build supports it: Vulkan (AMD), CUDA (NVIDIA) or Metal (macOS). A Vulkan loader alone does not establish GPU support in the installed CLI.

Installing it

Binaries and models are user-managed; VoiceStudio does not download them automatically or offer them through Model Catalogue.

  1. Install or build whisper-cli (whisper-cli.exe on Windows) with the GPU backend you want.
  2. Download a compatible Whisper GGML model, for example ggml-base.bin.
  3. On Linux and macOS, give the CLI execute permission: chmod +x /path/to/whisper-cli.

Selecting it

Set these environment variables before starting VoiceStudio (all desktop OSes):

  • OMNIVOICE_ASR_BACKEND=whisper-cpp selects the engine explicitly.
  • OMNIVOICE_WHISPER_CPP_BIN points to whisper-cli.
  • OMNIVOICE_WHISPER_CPP_MODEL points to your downloaded model file.
  • OMNIVOICE_WHISPER_CPP_DEVICE optionally selects the native GPU index.
  • OMNIVOICE_WHISPER_CPP_BACKEND=cpu disables native GPU acceleration.

Automatic ASR selection is unchanged: whisper.cpp is used only when you pick it.

Where it looks

  • CLI: backend/engines/whisper_cpp/bin/ or PATH; OMNIVOICE_WHISPER_CPP_DIR selects its directory. Legacy main executables are accepted only in an explicitly selected or package-local directory.
  • Models: backend/engines/whisper_cpp/models/, ~/.cache/whisper/, or the existing Hugging Face cache for ggerganov/whisper.cpp. Only a bare model name (such as ggml-base.bin) is searched there; a path in OMNIVOICE_WHISPER_CPP_MODEL that does not exist is reported as missing instead of loading a same-named model from another folder. This also applies to explicitly selected current-directory paths such as ./ggml-base.bin or .\ggml-base.bin; use the bare name for a local cache search.

Source language and translation

  • If the CLI omits detected-language metadata, or returns malformed metadata, the language stays unknown instead of defaulting to English. In interactive dubbing, select a source language before translating. Batch dubbing needs a language: when ASR reports none, the job stops with a message asking you to pick the source language.
  • The source language you choose in Dubbing applies everywhere the source text is read: the translation request, the Argos language packs checked and installed for the job, and the glossary terms extracted from it. Auto-detection only fills in while no explicit choice is made.
  • Translation requires a multilingual model; English-only .en, Turbo and Distil checkpoints are rejected for translation. That decision follows the checkpoint's own file name, so a model cached in the Hugging Face hub (stored as a symlink into a content-addressed blob) is judged by the cached entry's name, not the blob's hash.

Licences

The licence check follows the file rather than the name: a selected model keeps the terms of the repository that physically holds it, even when a cache entry points somewhere else. Models resolved from a Hugging Face cache retain their source repository licence check. Other custom local files have no assumed repository identity; their terms remain your responsibility.

Quirks

  • Global CPU selection always disables GPU use.
  • Diagnostics use conservative CPU estimates because the CLI's GPU providers are not verified.
  • Older CLI builds without full JSON output can transcribe without word timings.
  • Whole-file recognition follows the shared duration-based deadline, including time spent converting audio; an older CLI's JSON fallback spends the remaining time. See timeout settings.
  • Audio conversion permits local file and pipe inputs only. Remote URL entries in uploaded playlists are rejected.