WAVE-7B
by tsinghua-ee
Unified audio-visual embeddings for text, audio, silent video, and synchronized clips
tsinghua-ee/WAVE-7Bmixpeek://audio_extractor@v1/tsinghua_wave_7b_v1Overview
WAVE 7B is a Qwen2.5-Omni based embedding model for unified audio-visual retrieval. It creates a shared representation space for text, audio, silent video, and synchronized audio-video inputs, with prompt-aware embeddings for instruction-specific retrieval.
On Mixpeek, WAVE is a strong candidate when agents need to search multimodal observations where sound and motion both matter. A support agent can retrieve the clip where a machine squeals before stopping; a media agent can find a scene by its crowd sound and camera motion; an inspection agent can search for audiovisual anomalies without relying on transcripts alone.
Architecture
7B-class multimodal embedding model built on Qwen2.5-Omni with hierarchical feature fusion and a dual audio encoder for speech and environmental sound. It is trained with multimodal, multitask contrastive objectives across text, audio, video, and audio-video pairs.
Mixpeek SDK Integration
import { Mixpeek } from "mixpeek";const mx = new Mixpeek({ apiKey: "API_KEY" });await mx.collections.ingest({collection_id: "field-observations",source: { url: "s3://field-video/" },feature_extractors: [{feature: "audio_embeddings",model: "tsinghua-ee/WAVE-7B"}]});
Capabilities
- Any-to-any retrieval across text, audio, video, and audio-video clips
- Prompt-aware embeddings for task-specific search
- Strong audio and audiovisual retrieval performance
- Apache 2.0 license
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| MMEB-v2-video | Overall | 59.9 | WAVE model card |
| AudioCaps | Audio retrieval | 44.2 | WAVE model card |
| VGGSound | Audio-video retrieval | 25.0 | WAVE model card |
Performance
Specification
Research Paper
WAVE: Learning Unified and Versatile Audio-Visual Embeddings with Multimodal LLM
arxiv.orgBuild a pipeline with WAVE-7B
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Open Studio