End-to-End VITS Outperforms Two-Stage Tacotron 2 Synthesis for Kazakh: A Comparative Study on a New Multi-Domain Speech Corpus

Text-to-speech (TTS) synthesis for low-resource languages such as Kazakh is still difficult: the language is agglutinative, uses an extended Cyrillic alphabet and has few annotated speech corpora. We present Kazakh TTS systems trained on a new corpus of about 3000 pages of educational and scientific text from 23 domains, read by two professional speakers. A character-level Tacotron 2 model with guided attention and a Parallel WaveGAN vocoder was trained in ESPnet2 on the female voice, and end-to-end VITS models were trained at 24 kHz for both voices. Guided attention produced a sharp monotonic alignment within 21 epochs. On a new test set of 138 sentences, transcribed with a Kazakh-specific automatic speech recognition (ASR) model, VITS reached a character error rate (CER) of 1.94% (male) and 2.08% (female), against 8.79% for Tacotron 2, whose attention failed on 6.2% of the utterances. On 60 sentences with formulas, numbers and abbreviations, normalised by the system’s text front-end, Tacotron 2 failed on 11.7% of the utterances, while VITS had no failures and lowered the CER from 19.7% to 7.2%. The multilingual Whisper model reported about ten times higher error rates on the same audio, which shows that Whisper is not a reliable intelligibility measure for Kazakh. In a listening test with seven native speakers, Tacotron 2 with Parallel WaveGAN obtained a mean opinion score of 4.5 out of 5 for both naturalness and intelligibility.