Top AI Product

SwanVoice is ByteDance’s bid at expressive long-form zero-shot speech synthesis — a model that can speak in someone’s voice it has never heard before, sustain that voice across minutes of audio, and do it convincingly across both monologue and dialogue. The combination of “long-form” and “expressive” is the hard part most TTS systems quietly skip.

## Monologue and dialogue together

Most zero-shot voice cloning ships as a monologue tool: feed it a few seconds of reference audio and a paragraph of text and you get a single speaker reading flatly. SwanVoice targets dialogue too — turn-taking with consistent identity per speaker, including the prosodic shifts a real conversation has between question, answer, and aside. That distinction matters for the actual product surfaces: audiobooks, video dubbing, character voices in games, agent voices in long calls.

## Why “long-form” is the hard requirement

Long-form is where most TTS embarrasses itself. Voice identity drifts, expressive range collapses to a single tone, and pacing flattens. Solving that requires more than higher sample quality — it needs sustained control over prosody and identity across minutes of audio without the model forgetting whose voice it’s supposed to be. SwanVoice is part of ByteDance’s SwanAIGC audio research effort, alongside the SwanBench-Speech benchmark that explicitly evaluates exactly these long-form failure modes.

## Why it matters

Voice agents and generative video both need voice that holds up at length, not just sounds good for five seconds. Models trained and benchmarked specifically for long-form, multi-speaker, expressive synthesis are what unlocks the next layer of voice products.