Russian speech corpus improves long continuous text to speech systems
Balalaika-Longform: A Russian Speech Corpus for Continuous Long-Form Text-to-Speech
Sound
Summary
Generating speech for long texts is challenging because errors like missing words or early stopping are hidden in short sentences. The authors created Balalaika-Longform, a large collection of Russian speech recordings lasting up to 15 minutes, to help train and test speech systems on longer passages. They experimented with different training lengths and showed that using longer audio segments for training reduces errors in speech synthesis for long texts. This database helps better understand how training duration affects continuous speech generation.
What this means in practice
- •For speech technology developers: Create and improve long-form Russian text-to-speech models using a large continuous speech dataset for realistic training and evaluation.
- •For audiobook producers: Use models fine-tuned on long-form speech to generate more accurate, continuous Russian audiobook narrations with fewer speech errors.
Authors
Nikita Vasiliev, Kirill Borodin, Vasilii Kudryavtsev, Maxim Maslov, Grach Mkrtchian
Abstract
Long-form text-to-speech must retain requested words over extended generations, yet sentence-level training and evaluation can hide omissions and early stops. We introduce Balalaika-Longform, an open Russian corpus of 189 hours in continuous units of 30 seconds to 15 minutes. Long units and matched short windows support fine-tuning comparisons on the same source recordings, and the accompanying evaluation retains every synthesis attempt. We fine-tune CosyVoice3, Qwen3-TTS, VoxCPM2 and F5-TTS on either view and test continuous synthesis on 50 text-voice pairs with voices unseen in fine-tuning, at about 75, 300 and 1,200 words. At 1,200 words, CosyVoice3 WER falls from 99.9% after short-window fine-tuning to 47.3% with long targets, and to 16.6% with punctuated, format-matched transcripts; paired intervals favor long targets for Qwen3-TTS and VoxCPM2, and by a small margin for F5-TTS, whose absolute WER stays above 90%. The results isolate the effect of training sequence length under a fixed continuous-generation protocol.