wav2vec2-bert-base-s-emotion-russian-resd

wav2vec2-bert-base-s-emotion-russian-resd

Speech emotion recognition for Russian over seven classes: anger, disgust, enthusiasm, fear, happiness, neutral, sadness.

Audio tower jonatasgrosman/wav2vec2-large-xlsr-53-russian (WavLM/wav2vec2 fine-tuned for Russian ASR) paired with the text encoder DeepPavlov/rubert-base-cased; the model reads the waveform and a transcript together. Trained on Aniemore/resd.

• Parameters: 493M · weights: 1887 MiB (fp32)
• Input: 16 kHz mono waveform and the utterance transcript
• Quantized builds: wav2vec2-bert-base-s-emotion-russian-resd-quantized — INT8, FP8 and INT4, up to 5.8x smaller on disk at the same score

Results

<img src="assets/panel.svg" alt="macro-F1 across three test sets" width="720">

The last row is the audio branch alone. Production rarely has a gold transcript, and the gap between 0.718 with one and 0.702 without is what a transcript is worth here.

<img src="assets/confusion.svg" alt="Confusion matrix on RESD test" width="640">

Usage

import torch, librosa
from transformers import AutoConfig, AutoTokenizer, AutoModel

repo = "Aniemore/wav2vec2-bert-base-s-emotion-russian-resd"
cfg = AutoConfig.from_pretrained(repo, trust_remote_code=True)
model = AutoModel.from_pretrained(repo, trust_remote_code=True).eval()
tok = AutoTokenizer.from_pretrained(repo)

# Resample to 16 kHz. RESD ships at 44.1 kHz, and 44.1 kHz audio
# labelled as 16 kHz is stretched 2.8x in time — the model answers,
# it just answers about other audio.
wav, _ = librosa.load("clip.wav", sr=16000, mono=True)
wav = torch.tensor((wav - wav.mean()) / (wav.std() + 1e-7))[None]
t = tok("расшифровка реплики", return_tensors="pt")
with torch.no_grad():
    logits = model(input_ids=t['input_ids'], input_values=wav,
                   text_attention_mask=t['attention_mask'],
                   audio_attention_mask=torch.ones_like(wav).long()).logits

probs = logits.softmax(-1)[0]
print({model.config.id2label[i]: round(p.item(), 3) for i, p in enumerate(probs)})

<details><summary>Loading the audio without librosa</summary>

# torchaudio
import torchaudio
wav, sr = torchaudio.load("clip.wav")
wav = torchaudio.functional.resample(wav, sr, 16000).mean(0).numpy()

# torchcodec, the newer decoder
from torchcodec.decoders import AudioDecoder
wav = AudioDecoder("clip.wav", sample_rate=16000).get_all_samples().data.mean(0).numpy()

# straight from the dataset — `datasets` resamples on the column, so
# the mixed 16/44.1 kHz in RESD is handled for you
from datasets import load_dataset, Audio
ds = load_dataset("Aniemore/resd", split="test")
ds = ds.cast_column("speech", Audio(sampling_rate=16000))
wav = ds[0]["speech"]["array"]

</details>

Evaluation protocol

Audio resampled to 16 kHz mono, clips capped at 12 s, normalized per utterance, padding masked. UA is macro-averaged recall, WA is accuracy, F1 is macro-averaged. All three test sets went through the same harness, so the rows are comparable to each other.

The RESD split matches fold 1 of EmoBox bit for bit. The top entry there is WavLM-large at WA 56.47 / UA 55.87 / F1 55.82. These numbers are higher, but the training protocol differs — EmoBox freezes the encoder and trains a probe, this is a full fine-tune — so read it as a different recipe on the same split, not as a like-for-like win.

Limitations

RESD is acted and class-balanced. Real speech is neither. The Dusha podcast row above is the honest signal for spontaneous audio, and it is far below the RESD row. Spontaneous Russian is roughly 93% neutral, and a model tuned on balanced acted data over-predicts the emotional classes on it. Measure on your own material, and calibrate the neutral logit if you deploy.

Seven classes only, Russian only, single-speaker clips.

Citation

@misc{aniemore,
  author = {Lubenets, Ilya and Davidchuk, Nikita and Amentes, Aleksandr},
  title  = {Aniemore: an open library for emotion recognition in Russian speech},
  url    = {https://github.com/Aniemore/Aniemore},
  year   = {2023}
}

License

MIT.