# QwenAudio/ThinkSound

[NeurIPS 2025] PyTorch implementation of [ThinkSound], a unified framework for generating audio from any modality, guided by Chain-of-Thought (CoT) reasoning.

Repository: https://github.com/QwenAudio/ThinkSound
Canonical: https://ross.abutalabs.com/products/thinksound
Homepage: https://thinksound-project.github.io/
Language: Python
License Family: other
Topics: aigc-audio, foley-sound-synthesis, text-to-audio, text-video-to-audio, tta, video-to-audio
Last push: 2026-04-03T03:07:58+00:00

## Health v2 (maintenance only)
Score: 52/100 (v2, computed 2026-09-03T02:39:23.370411+00:00)
- activity 75, release rhythm 35, longevity 30
- inputs: {"age_days": 433, "days_push": 152, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1378, forks 82 (observed 2026-08-28T04:04:33.625378+00:00)

## What it is
ThinkSound is a PyTorch implementation of a NeurIPS 2025 framework that generates and edits audio from video, text, or audio inputs using Chain-of-Thought reasoning from multimodal large language models to guide a flow-matching audio foundation model. It also includes the AudioCoT dataset and hosts the follow-up PrismAudio project on a separate branch.

## Use cases
- generate foley sound effects for videos
- convert video into audio soundscapes
- generate audio from text descriptions
- edit audio in a video with natural language instructions
- add soundtracks to AI-generated videos
- research chain-of-thought reasoning for audio generation
- train or fine-tune video-to-audio models

## When to choose
- you need state-of-the-art video-to-audio or any-to-audio generation with reasoning guidance
- you want interactive, object-centric audio refinement for video content
- you are doing research on CoT-guided audio synthesis and need the AudioCoT dataset

## When to avoid
- you need a production-ready, licensed product with no research code caveats (no license is specified)
- you lack GPU resources for large multimodal model inference
- you need simple text-to-speech rather than generative foley/sound effects

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, audio-processing, llm-inference, deep-learning
- domain: artificial-intelligence, machine-learning
- platform: python
- tags: text-to-audio, video-to-audio, foley-sound-synthesis, chain-of-thought, flow-matching, aigc-audio, pytorch, multimodal, audio-editing, research-code, audio, video, natural-language-processing, gpu, linux

## Member repositories
- QwenAudio/ThinkSound (main) score 52

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:33.625378+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:40:21.072221+00:00, confidence not recorded.
  - readme: https://github.com/QwenAudio/ThinkSound (fetched 2026-08-28T04:04:33.625378+00:00, sha 42e4fc2eba56)
  - homepage: https://thinksound-project.github.io/ (fetched 2026-08-29T11:56:26.341313+00:00, sha 2ad582054d06)
- Data as of 2026-08-30T08:39:29.467469+00:00.