Models and media tools¶
The interface, ASR backend, weights, and media tools are separate. Prepare them before a long recording. The validated primary workflow uses Chinese podcasts on Apple Silicon, MLX Whisper, and large-v3-turbo.
Match the backend¶
| Computer | Processing path | Model directory |
|---|---|---|
| Apple Silicon Mac | MLX Whisper; validated Chinese workflow | model/whisper-large-v3-turbo/ |
| Windows | faster-whisper/CTranslate2; separate runtime setup | model/faster-whisper-large-v3-turbo/ |
The default root is ~/Documents/SpeechToTextData/, under the user Documents folder on Windows. Weights belong only in its model/ directory. Repository Downloads/ is temporary reference material, not normal model resolution.
Installers and the pipeline do not download weights. Prepare the exact files manually:
macOS / MLX Whisper¶
Source: mlx-community/whisper-large-v3-turbo. Keep these files together:
| Target under the data root | Manual download |
|---|---|
model/whisper-large-v3-turbo/config.json |
config.json |
model/whisper-large-v3-turbo/weights.safetensors |
weights.safetensors |
Do not mix the other backend's format or put weights in the source repository.
Windows / faster-whisper¶
Source: h2oai/faster-whisper-large-v3-turbo. Keep these files together:
| Target under the data root | Manual download |
|---|---|
model/faster-whisper-large-v3-turbo/config.json |
config.json |
model/faster-whisper-large-v3-turbo/model.bin |
model.bin |
model/faster-whisper-large-v3-turbo/preprocessor_config.json |
preprocessor_config.json |
model/faster-whisper-large-v3-turbo/tokenizer.json |
tokenizer.json |
model/faster-whisper-large-v3-turbo/vocabulary.json |
vocabulary.json |
Do not mix the other backend's format or put weights in the source repository.
FFmpeg and ffprobe¶
Both must be discoverable by the CLI doing the work. macOS can use Homebrew:
brew install ffmpeg
ffmpeg -version
ffprobe -version
Windows can use an existing PATH installation or vcpkg:
vcpkg install "ffmpeg[ffmpeg,ffprobe]:x64-windows"
ffmpeg -version
ffprobe -version
Qt Settings can expose the real local vcpkg root. Do not copy another computer's path. Bundled Qt libraries are not media tools.
Check before ASR¶
atot models
atot doctor
atot config set local.asr_model large-v3-turbo
Models lists aliases/files; doctor checks the runtime too. A rename cannot turn npz into safetensors or model.bin into MLX weights. Model changes affect future tasks, not existing published text.
Resources and candidates¶
Start with the default Chinese workflow. Speed/memory vary with hardware, duration, and backend; there is no universal real-time-factor promise.
Windows CUDA/cuDNN must satisfy the actual CTranslate2 wheel. CUDA 13 alone does not supply explicitly requested CUDA 12 DLLs. Follow the error or select CPU explicitly.
Inventory entries include smaller and experimental models. Listing is not promotion, and language choices are not equal validation evidence. Keep research separate from first production setup.
Moving or updating¶
Keep weights out of Git. Move model/ with the data root, then verify through the CLI launched against that root. Clearing temporary caches should not delete your weights.
The inventory defines aliases, exact files, and manual URLs.