Small-E: Small Language Model with Linear Attention for Efficient Speech Synthesis
Théodor Lemerle, Nicolas Obin, Axel Roebel
TL;DR
This work tackles the inefficiency of decoder-only transformers in text-to-speech by introducing Linear Causal Language Model blocks and Position-Aware Cross-Attention, enabling efficient training on long sequences (up to ~30s) with a compact 64M-parameter model. By framing TTS as conditional codec language modeling with RVQ and leveraging cross-modal PACA, Small-E achieves competitive zero-shot voice cloning while dramatically reducing training cost ($O(N)$) compared to standard self-attention. Empirical results show faster training throughput (≈62% gain) and improved or comparable objective/perceptual quality versus baselines like YourTTS, with ablations confirming PACA’s effectiveness in reducing skips and repetitions. The approach demonstrates that small, efficiently designed LM-based TTS systems can deliver high-quality synthesized speech and sets the stage for streaming and embedded TTS applications.
Abstract
Recent advancements in text-to-speech (TTS) powered by language models have showcased remarkable capabilities in achieving naturalness and zero-shot voice cloning. Notably, the decoder-only transformer is the prominent architecture in this domain. However, transformers face challenges stemming from their quadratic complexity in sequence length, impeding training on lengthy sequences and resource-constrained hardware. Moreover they lack specific inductive bias with regards to the monotonic nature of TTS alignments. In response, we propose to replace transformers with emerging recurrent architectures and introduce specialized cross-attention mechanisms for reducing repeating and skipping issues. Consequently our architecture can be efficiently trained on long samples and achieve state-of-the-art zero-shot voice cloning against baselines of comparable size. Our implementation and demos are available at https://github.com/theodorblackbird/lina-speech.