CALM and Pocket TTS: continuous-latent audio modeling and compact CPU distillation

What changed

Conditions a consistency-model sampler on transformer context to generate continuous audio-VAE frames; the expanded2026account distills a313M teacher into a six-layer Pocket TTS backbone while retaining its sampling head.

Scientific scope

Conditions a consistency-model sampler on transformer context to generate continuous audio-VAE frames; the expanded2026account distills a313M teacher into a six-layer Pocket TTS backbone while retaining its sampling head. Original2025method and expanded2026PocketTTS account bundled, not two achievements. Inherits MAR, consistency models and later Lagrangian self-distillation; audioVAE remains lossy and temperature heuristic. Expanded paper uses90Mgenerator plus decoder for advertised100M, not the whole20MVAE. CPU results are hardware-specific, not WRS reproduction; LibriSpeech evaluation uses enhanced audio for all compared baselines, and lowest WER is tied. First2equal in v3; later multilingual/training-code releases not backdated.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Original2025method and expanded2026PocketTTS account bundled, not two achievements. Inherits MAR, consistency models and later Lagrangian self-distillation; audioVAE remains lossy and temperature heuristic. Expanded paper uses90Mgenerator plus decoder for advertised100M, not the whole20MVAE. CPU results are hardware-specific, not WRS reproduction; LibriSpeech evaluation uses enhanced audio for all compared baselines, and lowest WER is tied. First2equal in v3; later multilingual/training-code releases not backdated.

Contributors

• Source-listed authors and technical contributors · SCIENCE: Source-listed authors or explicitly credited technical contributors for the cited versions. Family entries retain source-specific attribution; a combined list is not one original-paper byline or equal individual work. Names alone do not establish resolved identities.

• Continuous Audio Language Models
• Kyutai scientific publications index
• Continuous Audio Language Models — expanded Pocket TTS account

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction