Modelpublic

Aniemore/hubert-emotion-v1-crosslingual

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face

0likes20downloads

Model Card

hubert-emotion-v1-crosslingual

Speech emotion recognition over seven classes — anger, disgust, enthusiasm, fear, happiness, neutral, sadness.

Same architecture as `Aniemore/hubert-emotion-russian-resd`, retrained on a mix of 27,939 clips spanning eight languages and three speaking registers instead of one acted Russian corpus. The point of the change is spontaneous speech: the previous release was trained only on acted dialogue, where every class is equally frequent and every utterance is performed, and real speech is neither.

Quantized builds ship in the same repository under int8/, fp8/ and int4/.

Results

test set	what it is	macro-F1	UA	WA	previous release
RESD test	acted Russian, 7 balanced classes	0.6167	0.6158	0.6214	0.7522
Dusha podcast test	spontaneous Russian, majority neutral	0.3147	0.5898	0.5942	0.1477
CAMEO test	7 non-Russian languages	0.5293	0.5363	0.6350	0.2404

Read the first two rows together. The acted score goes down and the spontaneous score goes up; both follow from the same change, and which one matters is a deployment question. If your audio is read or performed speech, the previous release may still suit you better.

<details><summary>About the CAMEO row</summary>

CAMEO ships no train/test partition, and the usual way to make one — a random split over clips — puts nearly every test speaker into training as well: three of its twelve constituent corpora contain a single speaker each, so no clip-level split of them can be speaker-disjoint even in principle. The number above is reported for completeness. Treat it as an in-domain figure, not as evidence of cross-lingual transfer.

</details>

<details> <summary><b>Per class, across the panel</b> — recall and F1 for every class on every test set</summary>

Each card leads with the macro-F1 for that set; the rows are the detail behind it. Both per-class numbers are shown because they disagree in a way that matters: recall rewards a class the model over-predicts, so on the spontaneous set the minority classes reach decent recall at poor F1 — most clips called sad there are not sad. If you are going to act on one class, read its F1.

Each set keeps its own class list: the spontaneous corpus has four classes and the other two have seven, and there is no correspondence between positive and any single one of happiness/enthusiasm to line them up with.

</details>

Per-class recall on spontaneous speech

class	this model	previous release
`angry`	0.6287	0.5329
`neutral`	0.5848	0.1601
`positive`	0.8445	0.5854
`sad`	0.3010	0.2136

neutral carries most of real speech and is the class the previous release missed.

Quantized variants

subfolder	scheme	weights	vs fp32	macro-F1	UA	WA
(root)	fp32	1206 MiB	1.0x	0.6167	0.6158	0.6214
`int8`	W8A16	351 MiB	3.4x smaller	0.6195	0.6190	0.6250
`fp8`	W8A16-float	343 MiB	3.5x smaller	0.6190	0.6194	0.6250
`int4`	W4A16_ASYM	209 MiB	5.8x smaller	0.6194	0.6204	0.6250

Weight-only, round-to-nearest, no calibration. Every variant lands within the seed spread of the fp32 parent on RESD test, so the choice is about download size rather than about quality.

Training data

corpus	clips	language	register
RESD	948	Russian	acted dialogue
Dusha crowd	6,800	Russian	acted, crowd-sourced
CAMEO	6,800	7 languages	12 corpora, no Russian
Dusha podcast	6,060	Russian	spontaneous podcast speech
IEMOCAP	4,735	English	elicited dyadic sessions
ASVP-ESD	2,596	multilingual	mixed register
total	27,939	8 languages	3 registers

A slice is held out of every corpus in the mix, in the same proportion, and model selection is on that held-out split — never on any of the test sets above. Labels are unified to seven classes; four-class corpora are mapped upward and scored on the classes they actually contain.

Usage

python

import torch, librosa
from transformers import AutoModelForAudioClassification, AutoFeatureExtractor

repo = "Aniemore/hubert-emotion-v1-crosslingual"
model = AutoModelForAudioClassification.from_pretrained(repo).eval()
fe = AutoFeatureExtractor.from_pretrained(repo)

# Resample to 16 kHz. Do not skip it: RESD itself ships at 44.1 kHz,
# and handing the model 44.1 kHz audio while telling the extractor it
# is 16 kHz stretches time 2.8x and silently changes the answer.
wav, _ = librosa.load("clip.wav", sr=16000, mono=True)
x = fe(wav, sampling_rate=16000, return_tensors="pt", padding=True)
with torch.no_grad():
    probs = model(**x).logits.softmax(-1)[0]
print({model.config.id2label[i]: round(p.item(), 3) for i, p in enumerate(probs)})

For a quantized build, name the subfolder — only that subfolder is downloaded:

python

model = AutoModelForAudioClassification.from_pretrained(
    repo, subfolder="int8").eval()          # or "fp8", "int4"
fe = AutoFeatureExtractor.from_pretrained(repo, subfolder="int8")

Limitations

—Scores are the mean of two seeds; the seed spread on the 280-clip RESD split is ±0.03–0.05, so differences smaller than that are not differences.
—The spontaneous set has four classes where the model has seven, so its numbers are computed over a mapped label space and are not comparable to seven-class figures.
—Spontaneous scores are at zero decision bias. Calibrating the neutral threshold on your own development split will move them.
—This is a bimodal checkpoint: it takes a transcript alongside the audio. The transcripts used in training and evaluation are corpus-provided and clean, which ASR output is not.
—Inherited from `facebook/hubert-large-ls960-ft`; the licence follows the base model.