ThinkSound

Pioneering Audio Generation and Editing with Chain-of-Thought Reasoning

Abstract

While modern AI has made great strides in generating audio from video, creating high-fidelity sound that truly matches the nuance of visual content remains a major hurdle. Professional sound design requires complex reasoning about visual cues, acoustics, and timing—a skill that has been difficult to replicate in AI.

This paper introduces ThinkSound, a groundbreaking framework that teaches AI to "think" like a sound designer. By using Chain-of-Thought (CoT) reasoning, ThinkSound breaks down the complex task of audio generation into logical, manageable steps. This allows for not just creating sound from scratch but also for interactive, object-focused editing and refinement using simple natural language commands. To power this, we also present AudioCoT, a first-of-its-kind dataset designed to train models on this reasoning process. Our experiments show that ThinkSound sets a new state-of-the-art in both audio quality and relevance, performing exceptionally well even on complex, out-of-distribution movie scenes.

Synergy with Video Generation Models

ThinkSound seamlessly adds rich, synchronized soundscapes to videos created by leading generative models. The videos below were generated by their respective models; all audio was created by ThinkSound.

Veo + ThinkSound

Sora + ThinkSound

MovieGen + ThinkSound

V2A Comparisons on VGGSound (In-distribution)

Click on any thumbnail to load and play the video, comparing ThinkSound to other models.

CoT	Ground Truth	ThinkSound	MMAudio	See&Hear	V-AURA	Frieren	V2A-Mapper	FoleyCrafter
Playing Tennis Generate sounds of tennis hitting a racket and the ball bouncing...
Printer Printing Generate a continuous printer printing sound with periodic beeps...
Ripping Paper Start with a subtle tearing sound of paper being ripped...
Using Sewing Machines Generate ambient sewing room sounds with consistent sewing machine hum...
Playing Bongo Generate a lively percussion track featuring only rhythmic drum beats...
Chopping Food Generate rhythmic chopping sounds consistent with cutting meat or vegetables...
People Eating Crisps Generate audio focusing on clear, rhythmic chewing sounds...

V2A Comparisons on MovieGen Audio (Out-of-Distribution)

See how ThinkSound performs on challenging, out-of-distribution movie clips.

CoT	ThinkSound	Movie Gen Audio	MMAudio
Gentle Sucking Sounds Soft, steady background of light pacifier suckling...
Harmonious Strings Acoustic guitar strings humming and buzzing...
Old TV Humming Ambient background noise with faint static and white noise...
Intense Thunder A low wind hum and occasional crackles add to the stormy atmosphere...
High-Pitched Scraping High-pitched, sustained scraping sound of a tool on a metal rod...
Clattering Metal Keys Rhythmic sound of an old typewriter, focusing on the sharp metallic clatter...
Skateboard Grinding Steady rolling on a hard surface, with sharp scraping and grinding sounds...

Interactive Step-by-Step Foley Creation

V2A Gen → Object-Focus → Audio Inpainting

Generate a cheerful ukulele melody with light strumming and harmonious vocals from two young girls singing together.

Generated audio (paired with silent video):

Now, focus only on the singing and hand movements in the selected region.

Generated audio for the selected object:

Repair the masked (noisy) segment in this audio clip.

Repaired audio:

V2A Gen → Object-Focus → Audio Editing

Generate gentle wind sounds with consistent warbler chirping.

Generated audio (paired with silent video):

Focus on the bird, reduce the wind noise, and make the chirping crisp and clear.

Generated audio for the selected object:

Keep the warbler chirping and add an occasional robin call for contrast.

Edited audio:

Experiments

Main Results on VGGSound

ThinkSound outperforms all baselines across most objective metrics and all subjective metrics, achieving substantial improvements in audio quality and semantic alignment.

Table 1: Comparison of our ThinkSound foundation model...
Method	Objective Metrics						Subjective Metrics		Efficiency
	FD ↓	KL_PaSST ↓	KL_PaNNs ↓	DeSync ↓	CLAP_cap ↑	CLAP_CoT ↑	MOS-Q ↑	MOS-A ↑	Params	Time(s) ↓
GT	-	-	-	0.55	0.28	0.45	4.37±0.21	4.56±0.19	-	-
See&Hear	118.95	2.26	2.30	1.20	0.32	0.35	2.75±1.08	2.87±0.99	415M	19.42
V-AURA†	46.99	2.23	1.83	0.65	0.23	0.37	3.42±1.03	3.20±1.17	695M	14.00
FoleyCrafter	39.15	2.06	1.89	1.21	0.41	0.34	3.08±1.21	2.63±0.88	1.20B	3.84
Frieren†	74.96	2.55	2.64	1.00	0.37	0.34	3.27±1.11	2.95±1.09	159M	-
V2A-Mapper†	48.10	2.50	2.34	1.23	0.38	0.32	3.31±1.02	3.16±1.04	229M	-
MMAudio	43.26	1.65	1.40	0.44	0.31	0.40	3.84±0.89	3.97±0.82	1.03B	3.01
ThinkSound	34.56	1.52	1.32	0.46	0.33	0.46	4.02±0.73	4.18±0.79	1.30B	1.07
w/o CoT Reasoning	39.84	1.59	1.40	0.48	0.29	0.41	3.91±0.83	4.04±0.75	1.30B	0.98

Ablation Studies

We investigated the contribution of each component to validate the effectiveness of our design choices, focusing on text encoding and multi-modal integration.

Text Encoding Strategies

Table 2: Comparison of text encoder fusion strategies...
Method	FD ↓	KL_PaSST ↓	KL_PaNNs ↓	DeSync ↓	CLAP ↑
CLIP	39.84	1.59	1.40	0.48	0.41
T5 (CoT)	37.65	1.54	1.35	0.46	0.44
CLIP + T5	34.56	1.52	1.32	0.46	0.46

Multi-Modal Integration

Table 3: Comparison of multi-modal integration mechanisms
Integration	FD ↓	KL_PaSST ↓	KL_PaNNs ↓	DeSync ↓	CLAP ↑
audio only	37.13	1.58	1.37	0.50	0.43
linear video	38.96	1.58	1.38	0.46	0.45
gated video	34.56	1.52	1.32	0.46	0.46

Impact of Model Size

Table 4: Impact of model size results.
Size	FD ↓	KL_PaSST ↓	KL_PaNNs ↓	DeSync ↓	CLAP_CoT ↑
Small	40.80	1.64	1.38	0.46	0.41
Medium	36.80	1.56	1.34	0.46	0.44
Large	34.56	1.52	1.32	0.46	0.46

Frequently Asked Questions

What is ThinkSound?

How does Chain-of-Thought (CoT) help?

Chain-of-Thought allows the model to break down a complex task (like "create a soundtrack for this video") into smaller, logical steps. For example, it might first identify the main objects and actions, then reason about the environment's acoustics, and finally decide on the appropriate sounds and their timing. This step-by-step process leads to more accurate and contextually aware audio generation.

What makes ThinkSound different?

Can I try ThinkSound myself?