Model specs

Every place that takes a model, the CLI --model flag, get_backend(), the model form field on the HTTP API and the VAAK_MODEL environment variable, accepts the same string:

<backend>:<model>

<backend> is a registry key from vaak/backends/__init__.py.
<model> is everything after the first colon. Its meaning depends on the backend.

vaak models prints the backends available in the current environment.

Backends

`faster-whisper:<size>`

Whisper via CTranslate2. <size> is one of tiny, base, small, medium, large-v2, large-v3, large-v3-turbo, or distil-large-v3, or a local directory containing a CTranslate2 model. Multilingual; languages is None.

faster-whisper:small
faster-whisper:large-v3-turbo
faster-whisper:./checkpoints/whisper-small-hi-ct2

VAAK_COMPUTE_TYPE selects float16, int8_float16, int8 or float32. Extra: vaak[whisper].

`indic-conformer:<lang>`

AI4Bharat IndicConformer through NeMo. <lang> is an ISO 639-1 code for one of the supported Indian languages, for example hi, bn, ta, te, mr, gu, kn, ml, pa, or, as, ur. The backend maps the code to the matching AI4Bharat checkpoint. languages is {<lang>}.

indic-conformer:hi

Extra: vaak[nemo].

`parakeet:<hf-id>`

NVIDIA Parakeet TDT through NeMo. <hf-id> is a Hugging Face model id such as nvidia/parakeet-tdt-0.6b-v2. English only; languages is {"en"}.

parakeet:nvidia/parakeet-tdt-0.6b-v2

Extra: vaak[nemo].

`wav2vec2:<hf-id>`

Any CTC model loadable with transformers.AutoModelForCTC. <hf-id> is a Hub id or a local directory. languages is read from the model config when present, otherwise None.

wav2vec2:facebook/wav2vec2-base-960h
wav2vec2:ai4bharat/indicwav2vec-hindi

Extra: vaak[whisper] (it installs transformers and torch).

`vosk:<model-name>`

Vosk (Kaldi) models, CPU only. <model-name> is a name from the Vosk model list such as vosk-model-small-en-us-0.15 or vosk-model-small-hi-0.22, or a local directory. The backend downloads named models on first use. languages is derived from the model name.

vosk:vosk-model-small-hi-0.22

Extra: vaak[vosk].

`hf-whisper:<hf-id>`

Whisper through transformers, not CTranslate2. Slower than faster-whisper but loads any Hub checkpoint, including LoRA-merged fine-tunes from train/finetune_whisper_lora.py. <hf-id> is a Hub id or local directory. Multilingual; languages is None.

hf-whisper:openai/whisper-small
hf-whisper:<your-user>/whisper-small-hi-lora

Extra: vaak[whisper].

Parsing rules

Split on the first colon only. Hub ids and Windows paths can contain further colons.
Whitespace around the string is stripped.
An unknown backend prefix raises ValueError in Python and returns 404 from the HTTP API.
A missing colon raises ValueError. There is no default backend for a bare model name.