Tacit-TTS: From Autoregressive Decoding to Masked Prediction for Efficient Transcript-Free Voice Cloning
Jian Chen, You Zhang, Mark Vinton
Dolby Laboratories
Published September 29, 2026 · on the feed October 1, 2026 · ▲ 11 · arXiv 2609.38658
The abstract
TTS systems with autoregressive semantic modeling have demonstrated strong zero-shot voice cloning performance and rich expressive variation, but their sequential decoding incurs substantial latency. Non-autoregressive alternatives offer much faster generation, yet often rely on more restrictive reference conditioning, such as requiring transcripts of the reference speech during inference. We present Tacit-TTS, an efficient transcript-free zero-shot voice cloning system distilled from IndexTTS2. Our model replaces autoregressive text-to-semantic decoding with masked non-autoregressive generation, introduces training-free acoustic length estimation, and accelerates the flow-matching renderer through ReFlow distillation. Across two English and two Mandarin datasets, Tacit-TTS achieves competitive zero-shot quality while generating speech over 10x faster than IndexTTS2 for utterances longer than 5 seconds. Its transcript-free conditioning further supports cross-lingual and non-lexical references. We validate this capability using references from eight other languages, infant babble, and synthetic gibberish, where transcript-dependent systems often degrade or fail due to unreliable ASR transcripts.
The Brief
1:55A short podcast summary of the paper. Written and voiced by AI.
Transcript
Tacit-TTS. From Autoregressive Decoding to Masked Prediction for Efficient Transcript-Free Voice Cloning. By Jian Chen and two others, from Dolby Laboratories. Published on the twenty-ninth of September. No code repository is linked to it.
Speech synthesis systems that model semantics autoregressively are good at zero-shot voice cloning, with rich expressive variation. The price is latency, because their decoding runs one step after another. Non-autoregressive systems are much faster, but they often demand more of the reference: typically a transcript of the reference speech at inference time. That leaves a choice between speed and the freedom to clone from audio alone.
Tacit-TTS aims to have both. It is an efficient, transcript-free, zero-shot voice cloning system, distilled from IndexTTS2, that keeps the reference conditioning free of transcripts while dropping the sequential decoder.
Three changes do the work. Autoregressive text-to-semantic decoding is replaced with masked, non-autoregressive generation. Acoustic length is estimated without any training. And the flow-matching renderer is sped up through ReFlow distillation.
Testing covers two English datasets and two Mandarin ones. Tacit-TTS reaches competitive zero-shot quality while generating speech more than ten times faster than IndexTTS2, for utterances longer than five seconds. The transcript-free conditioning also opens up references that are cross-lingual or not lexical at all. To check this, the authors use references in eight other languages, infant babble, and synthetic gibberish. Those are cases where systems that depend on transcripts often degrade or fail, because the ASR transcripts they rely on are unreliable.
Babble and gibberish make the clearest test case here. A system that needs to know what the reference says has nothing to work with when the reference says nothing in particular. One that only listens to how it sounds can still clone the voice, and Tacit-TTS gets there at a fraction of its teacher's generation time.
Paper Quest · 7 chapters · about 40 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
A beautiful voice, delivered too slowly, and a faster rival that needs words you may not have.
You are building a system that speaks any sentence in the voice of a stranger, from a few seconds of their recorded speech. The best system you have sounds great, but it writes its speech plan one token after another, so long sentences are slow.
The road ahead
- 1 · The problem ~7 min Copying a stranger's voice well is slow when speech is planned one token at a time, and the fast alternatives need a transcript of the recording.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Fill in the whole speech plan in parallel over a few rounds, set its length from the reference's pace, and never need the reference's words.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min First size the plan from pace and syllables, then fill it in parallel, then turn it into sound with a sped-up renderer.
- 6 · The experiment ~6 min You’ll guess: For utterances longer than 5 seconds, how many times faster than its teacher IndexTTS2 does Tacit-TTS generate speech?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.