Classify Speech Emotion (emotion2vec)

@funaudiollm1 VersionClassify

Classifies speech affect in each AudioFrame with an emotion2vec model, emitting ranked [Classification] pairs or a single top-label String chosen by the connected output slot. Audio is resampled to 16 kHz internally for language-robust affect labelling.

How it fits

Typical backends

Call-quality routing where detected emotion gates a branch from an audio file.

Per-turn emotion analytics where each aggregated speech turn is scored and logged.

Caveats

State lifetimemodel_name, hub, granularity, and device are read once at startup; changing any of them requires a restart.
I/O contractAudio is resampled to 16 kHz internally regardless of the upstream sample rate, because the emotion2vec models are trained at that rate.
Hard constraintmodel_name must be the canonical id (the `iic/` prefix); a raw third-party repo path is not supported.
Parameter interactionhf_token is only consulted when hub is hf and the repo is gated or private.
Parameter interactiongranularity of frame scores per analysis frame rather than over the whole clip; most affect routing wants the utterance default.
FallbackAn empty AudioFrame returns an empty result without running the model; downstream must handle the zero-element or empty-string case.

Versions

84d1b868latestdefault
Automated release
6/28/2026