VoiceCLAP-Small-v2

Voice-text contrastive (CLAP-style) embedding model — the successor to laion/voiceclap-small, trained with emotion-led MOSS-Audio short captions. Better than v1 on every benchmark we measure, at identical size and inference cost.

Same dual-tower architecture as v1: a BUD-E-Whisper_V1.1 audio encoder paired with sentence-transformers/all-MiniLM-L6-v2 on the text side, joined by an MLP projection on each side and trained with the SigLIP sigmoid contrastive loss.

Architecture	dual-tower CLAP (BUD-E-Whisper-Small + MiniLM-L6-v2)
Audio encoder	Whisper-style: 12 layers × 768 dim × 12 heads, 80-mel input @ 16 kHz
Text encoder	BERT/MiniLM, 6 layers × 384 dim, mean-pooled
Joint embedding	768-d, L2-normalised
Loss	SigLIP (sigmoid contrastive)
Total parameters	~110 M
Training	40 M samples (20 epochs × 2 M), best checkpoint epoch 19

What's new vs v1

v1 sampled k=2 uniformly-chosen MOSS-Audio attribute sentences per clip as the caption. v2 replaces this with an emotion-led short caption: the MOSS-Audio-8B-Thinking EMO sentence (a direct natural-language description of the emotional state) plus one randomly sampled talking-style sentence, re-drawn every epoch. Captions stay 50/50 blended with each corpus's original captions. The emotion-first structure concentrates contrastive signal on the emotion subspace without sacrificing style coverage.

Evaluation

Benchmark	v1 (released)	v2 (this model)	Δ
EmoNet-Voice top-1	0.0902	0.1015	+13% rel
EmoNet-Voice Spearman ρ	0.2280	0.2561	+12% rel
MAEB-voice mean (8 tasks)	0.3861	0.3893	+0.8%

The ρ gain also clears every arm of the v1 caption-sampling sweep (best: 0.2399 at k=2). MAEB-voice shows no general-speech regression.

Training data

Trained on the open 9-corpus mixture used in the VoiceNet paper:

• emolia-balanced-5M-subset (annotated subset of Emilia)
• laions_got_talent_clean_with_captions
• majestrino-data
• synthetic_vocal_bursts + improved_synthetic_vocal_bursts
• ears, expresso, voxceleb1, voxceleb2 (FCaps captions)

MOSS-Audio-8B-Thinking annotations (18 prompt groups, 61 attribute values per clip) provide the EMO + style sentences for the three large corpora.

Usage

import torch
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("laion/voiceclap-small-v2", trust_remote_code=True).eval()
tok = AutoTokenizer.from_pretrained("laion/voiceclap-small-v2")

# audio: raw mono waveform at 16 kHz
import soundfile as sf
wav, sr = sf.read("clip.wav", dtype="float32")
audio_emb = model.encode_waveform(torch.from_numpy(wav))

# text
t = tok(["a person speaking with quiet pride in their voice"], padding=True, return_tensors="pt")
text_emb = model.encode_text(t["input_ids"], attention_mask=t["attention_mask"])

score = (audio_emb @ text_emb.T).item()

Conversion from the training checkpoint was verified functionally against the original open_clip implementation (cosine ≥ 0.9999 on both towers).

Sibling models

• laion/voiceclap-large-v2 — 7B single-tower successor trained with Prototypical Contrastive loss
• laion/voiceclap-small, laion/voiceclap-large — v1 releases

License

cc-by-nc-4.0