Can SenseVoiceSmall be registered and run locally instead of downloading from HuggingFace?

[sumedh-garimella]

I am trying to run a local version of SenseVoiceSmall (I am assuming that my production environment will not have access to download a model from HuggingFace), but am coming across the error '{model path} is not registered', despite attempting to use the tables.register() method. Is it possible to register a local copy of the SenseVoiceSmall and use it with AutoModel, and if so, how do I do so? Thanks in advance!

Reference code:
model_dir = "SenseVoiceSmall/model.pt" # downloaded from HuggingFace to be used locally
tables.register('SenseVoiceSmall', 'SenseVoice/model.py') # SenseVoice is the cloned SenseVoice code
#tables.register('SenseVoiceSmall', model_dir)
model = AutoModel(model=model_dir, device="cuda:0", hub=None)

[LauraGPT]

model_dir should be absolute path.

[LauraGPT]

Yes. You can run SenseVoiceSmall fully offline, but model must point to the complete model directory, not to model.pt.

The directory needs the files referenced by its config, including at least:

SenseVoiceSmall/
├── config.yaml
├── model.pt
├── tokens.json
├── am.mvn
└── chn_jpn_yue_eng_ko_spectok.bpe.model

Use an absolute directory path and disable both update checks:

from funasr import AutoModel

model_dir = "/absolute/path/to/SenseVoiceSmall"

model = AutoModel(
    model=model_dir,
    hub="ms",
    device="cuda:0",
    disable_update=True,
    check_latest=False,
)

result = model.generate(
    input="/absolute/path/to/audio.wav",
    language="auto",
    use_itn=True,
)
print(result)

You do not need tables.register(). That registry is for mapping a model class name to Python implementation code; it is not how checkpoint files are registered.

The original model="SenseVoiceSmall/model.pt" fails because FunASR expects to resolve config.yaml, tokenizer files, frontend files, and model.pt relative to a model directory. Copy or download the whole snapshot on a connected machine, then transfer that directory to the offline host.

I verified the command against current FunASR main with all network requests forced to an unreachable proxy: the local SenseVoiceSmall directory loaded as SenseVoiceSmall and transcribed the bundled English sample successfully. disable_update=True disables the FunASR version check; check_latest=False prevents the ModelScope freshness check for a local directory.

Relevant loader code: local ModelScope directory handling.