Text to speech models sustain clear voices during long talks
Speaking Longer Than Learned: Training-Free Long-Form Generation for Full-Stream Text-to-Speech
Sound
Summary
Text-to-speech models generate spoken words from written text, but they often struggle when making long speeches without extra training. The authors studied an existing model called VoXtream2 and found that its way of handling long talks can cause the voice to sound less accurate or less like the original speaker. They proposed two simple changes that don’t need more training to fix these problems, helping the model keep pronunciation clear and the speaker’s voice consistent even in long speeches.
What this means in practice
- •For voice assistant developers: Produce natural and consistent long speech outputs that maintain speaker identity without extra model training.
- •For audiobook producers: Generate lengthy narrated content with stable pronunciation and speaker voice for full books without retraining speech models.
Authors
Jeseung Lim, Jiseong Woo, Heeseung Kim
Abstract
Recent full-stream text-to-speech (TTS) models enable low-latency personalized synthesis as input text arrives. However, sustaining generation beyond the lengths seen during training without additional training remains largely unexplored. We systematically analyze VoXtream2, which extends generation by periodically discarding past context while retaining the acoustic prompt. We find that this strategy does not prevent degradation: pronunciation accuracy deteriorates near the end of the training range due to the range of positions in which the model operates rather than the loss of past context, while speaker similarity gradually declines throughout generation. Based on these findings, we propose two training-free modifications that reduce the attention window and strengthen attention to the acoustic prompt. Experiments show that these modifications preserve pronunciation accuracy and speaker similarity throughout long-form generation while retaining the naturalness of the original model. Ablations further show complementary effects of the two modifications.