Latent Speech-Text Transformer
Yen-Ju Lu, Yashesh Gaur, Wei Zhou, Benjamin Muller, Jesus Villalba, Najim Dehak, Luke Zettlemoyer, Gargi Ghosh, Mike Lewis, Srini Iyer, Duc Le
We introduce Latent Speech-Text Transformer, which patches long speech token sequences into latent units, improving text–speech transfer while cutting pre-training and inference compute, and significantly outperforming existing speech-text LLMs.
Abstract
Auto-regressive speech–text models pre-trained on interleaved text tokens and discretized speech tokens demonstrate strong speech understanding and generation, yet remain substantially less compute-efficient than text LLMs, partly due to the much longer sequences of speech tokens relative to text. This modality imbalance disproportionately allocates pre-training and inference compute to speech, potentially hindering effective cross-modal alignment and slowing performance scaling by orders of magnitude. We introduce the Latent Speech-Text Transformer (LST), which aggregates speech tokens into latent speech patches that serve as higher-level autoregressive units. This design aligns the sequence-modeling granularity between speech and text while improving computational efficiency. The resulting patches can align with textual units to facilitate cross-modal knowledge transfer and compactly capture recurring acoustic patterns such as silence. Across story-completion benchmarks under both compute-controlled and data-controlled settings, LST consistently improves speech accuracy while also improving text performance, achieving up to +6.5% absolute gain on speech HellaSwag in compute-controlled training (+5.3% in data-controlled training). Under compute-controlled scaling from 420M to 1.8B parameters in a near compute-optimal regime, gains grow with scale, and improvements persist up to 7B parameters under fixed-token budgets. These benefits extend to downstream tasks: LST stabilizes ASR adaptation and reduces the effective autoregressive sequence length during ASR and TTS inference, lowering computational cost without degrading reconstruction quality. The Code is available at https://github.com/facebookresearch/lst.
Aggregates speech tokens into latent patches for efficient speech-text modeling with cross-modal alignment.
- Proposes patch-based framework aggregating speech tokens into latent units
- Aligns sequence-modeling granularity between speech and text modalities
- Achieves up to +6.5% absolute gain on speech tasks under compute-controlled scaling
- Stabilizes ASR adaptation and reduces effective sequence length during inference
- Speech-text modeling
- Token aggregation
- Multimodal transformers
- LibriLight
- People's Speech
- Multilingual LibriSpeech
- Spotify
Focuses on half-duplex speech-text modeling; does not address full-duplex interaction
from the paperAnalysis restricted to pre-training stage without instruction fine-tuning or downstream adaptation
from the paperPatching strategies rely on forced alignments during pre-training; fully alignment-free approaches remain open
from the paperLimited to speech-text modality; not yet extended to image or video
from the paper
Authors did not state explicit future directions.
Author keywords
- Speech–Text Models
- Latent Patching
- Multimodal Alignment
- Large Language Models
Related orals
Multimodal Aligned Semantic Knowledge for Unpaired Image-text Matching
MASK aligns semantic knowledge between images and text using word embeddings as bridges to match out-of-distribution words in unpaired matching.
ScaleCUA: Scaling Open-Source Computer Use Agents with Cross-Platform Data
ScaleCUA scales open-source computer use agents with cross-platform dataset and dual-loop data pipeline.
VibeVoice: Expressive Podcast Generation with Next-Token Diffusion
Presents VibeVoice for zero-shot expressive long-form multi-speaker podcast generation using next-token diffusion.
UALM: Unified Audio Language Model for Understanding, Generation and Reasoning
UALM unified audio language model handles understanding, text-to-audio generation, and multimodal reasoning in single model with UALM-Reason for cross-modal generative reasoning.
MetaEmbed: Scaling Multimodal Retrieval at Test-Time with Flexible Late Interaction
MetaEmbed uses learnable meta tokens with matryoshka training to enable test-time scaling for multimodal retrieval balancing quality and efficiency.