Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders

• URL: http://arxiv.org/abs/2606.10029v1
• Date: Mon, 08 Jun 2026 18:09:37 GMT
• Title: Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders
• Abstract summary: We train BatchTopK sparse autoencoders on the LM backbone of CosyVoice3.<n>We label each feature from where it fires-text-aware context, 1-second speech clips, or both.<n>The recovered features are interpretable, spanning phonemes, laughter, accent prompts and speaker gender.
• Score: 8.27358250310327
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Language models increasingly serve as the backbone of text-to-speech (TTS) systems, yet we understand little about the representations they build when text and generated speech tokens share a single residual stream. We train BatchTopK sparse autoencoders on the LM backbone of CosyVoice3 and introduce a modality-aware auto-interp pipeline that labels each feature from where it fires-text-prefix context, 1-second speech clips, or both. The recovered features are interpretable, spanning phonemes, laughter, accent prompts and speaker gender. Steering through the SAE latent space shows these features are causal rather than merely descriptive: targeted interventions raise laughter probability from 0.02 to 0.79, flip perceived speaker gender, and control speech rate while preserving spoken content. SAE features thus serve both as interpretability objects and as control directions for TTS synthesis.

Related papers

• Shared Latent Representation for Joint Text-to-Audio-Visual Synthesis [57.5830191022097]
  A Text-to-Vec module generates Wav2Vec2 embeddings from text, which jointly condition speech and face generation.<n>We adopt a two-stage training: pretraining on Wav2Vec2 embeddings and finetuning on TTS outputs.<n>Experiments show that conditioning on TTS-predicted latent features outperforms cascaded pipelines.
  arXiv  Detail & Related papers  (2025-11-07T17:07:56Z)
• Listening or Reading? Evaluating Speech Awareness in Chain-of-Thought Speech-to-Text Translation [12.571782794778182]
  Chain-of-Thought (CoT) prompting has been introduced, with the expectation that jointly accessing speech and transcription will overcome these issues.<n>We find that it largely mirrors cascaded behavior, relying mainly on transcripts while barely leveraging speech.<n>Simple training interventions, such as adding Direct S2TT data or noisy transcript injection, enhance robustness and increase speech attribution.
  arXiv  Detail & Related papers  (2025-10-03T15:42:38Z)
• NVSpeech: An Integrated and Scalable Pipeline for Human-Like Speech Modeling with Paralinguistic Vocalizations [7.55995559331834]
  Paralinguistic vocalizations are integral to natural spoken communication.<n> NVSpeech bridges the recognition and synthesis of paralinguistic vocalizations.<n> NVSpeech offers the first open, large-scale, word-level annotated pipeline for expressive speech modeling in Mandarin.
  arXiv  Detail & Related papers  (2025-08-06T08:25:26Z)
• IndexTTS2: A Breakthrough in Emotionally Expressive and Duration-Controlled Auto-Regressive Zero-Shot Text-to-Speech [13.3878636941081]
  IndexTTS2 is a novel, general, and autoregressive model-friendly method for speech duration control.<n>It achieves disentanglement between emotional expression and speaker identity, enabling independent control over timbre and emotion.<n>It outperforms state-of-the-art zero-shot TTS models in terms of word error rate, speaker similarity, and emotional fidelity.
  arXiv  Detail & Related papers  (2025-06-23T08:33:40Z)
• DisfluencySpeech -- Single-Speaker Conversational Speech Dataset with Paralanguage [7.096838107088313]
  DisfluencySpeech is a studio-quality labeled English speech dataset with paralanguage. A single speaker recreates nearly 10 hours of expressive utterances from the Switchboard-1 Telephone Speech Corpus (Switchboard)
  arXiv  Detail & Related papers  (2024-06-13T05:23:22Z)
• Visual-Aware Text-to-Speech [101.89332968344102]
  We present a new visual-aware text-to-speech (VA-TTS) task to synthesize speech conditioned on both textual inputs and visual feedback of the listener in face-to-face communication. We devise a baseline model to fuse phoneme linguistic information and listener visual signals for speech synthesis.
  arXiv  Detail & Related papers  (2023-06-21T05:11:39Z)
• NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers [90.83782600932567]
  We develop NaturalSpeech 2, a TTS system that leverages a neural audio predictor with residual vectorizers to get the quantized latent vectors. We scale NaturalSpeech 2 to large-scale datasets with 44K hours of speech and singing data and evaluate its voice quality on unseen speakers. NaturalSpeech 2 outperforms previous TTS systems by a large margin in terms of prosody/timbre similarity, synthesis, and voice quality in a zero-shot setting.
  arXiv  Detail & Related papers  (2023-04-18T16:31:59Z)
• UnifySpeech: A Unified Framework for Zero-shot Text-to-Speech and Voice Conversion [63.346825713704625]
  Text-to-speech (TTS) and voice conversion (VC) are two different tasks aiming at generating high quality speaking voice according to different input modality. This paper proposes UnifySpeech, which brings TTS and VC into a unified framework for the first time.
  arXiv  Detail & Related papers  (2023-01-10T06:06:57Z)
• SpeechLM: Enhanced Speech Pre-Training with Unpaired Textual Data [100.46303484627045]
  We propose a cross-modal Speech and Language Model (SpeechLM) to align speech and text pre-training with a pre-defined unified representation. Specifically, we introduce two alternative discrete tokenizers to bridge the speech and text modalities. We evaluate SpeechLM on various spoken language processing tasks including speech recognition, speech translation, and universal representation evaluation framework SUPERB.
  arXiv  Detail & Related papers  (2022-09-30T09:12:10Z)
• Zero-Shot Text-to-Speech for Text-Based Insertion in Audio Narration [62.75234183218897]
  We propose a one-stage context-aware framework to generate natural and coherent target speech without any training data of the speaker. We generate the mel-spectrogram of the edited speech with a transformer-based decoder. It outperforms a recent zero-shot TTS engine by a large margin.
  arXiv  Detail & Related papers  (2021-09-12T04:17:53Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.