This paper substantiates the feasibility of using the text-to-text (TTT) approach instead of speech-to-text (STT) when creating parallel audio data, which is especially relevant for the tasks of machine translation of speech and the generation of multilingual audio corpora. The TTT method allows for the formation of high-quality parallel text pairs with their subsequent conversion to audio using text-to-speech (TTS) systems, thereby avoiding errors inherent in automatic speech recognition (ASR), such as distortions caused by noise, accents, and intonation variations. The study applies TTS and STT technologies to create a parallel Kazakh-Turkish speech base, showcasing the adaptability and versatility of the TTT approach in handling diverse linguistic and audio processing tasks. The conducted comparative analysis of Massively Multilingual Speech (MMS) and KazakhTTS models showed that KazakhTTS provides lower spectral distortion (MCD 154.6970) and better subjective sound quality (PESQ 1.1579) compared to MMS (MCD 162.4576, PESQ 1.0449). Despite lower intelligibility (STOI = 0.0444 vs. 0.2430 in MMS), KazakhTTS demonstrates high adaptability to the Kazakh language and is preferable for speech synthesis tasks within the framework of building a parallel corpus.

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Parallel Speech Corpus of Kazakh and Turkish Languages: Methodology of Creation and Application Prospects

  • Aidana Karibayeva,
  • Diana Rakhimova,
  • Madina Askapova

摘要

This paper substantiates the feasibility of using the text-to-text (TTT) approach instead of speech-to-text (STT) when creating parallel audio data, which is especially relevant for the tasks of machine translation of speech and the generation of multilingual audio corpora. The TTT method allows for the formation of high-quality parallel text pairs with their subsequent conversion to audio using text-to-speech (TTS) systems, thereby avoiding errors inherent in automatic speech recognition (ASR), such as distortions caused by noise, accents, and intonation variations. The study applies TTS and STT technologies to create a parallel Kazakh-Turkish speech base, showcasing the adaptability and versatility of the TTT approach in handling diverse linguistic and audio processing tasks. The conducted comparative analysis of Massively Multilingual Speech (MMS) and KazakhTTS models showed that KazakhTTS provides lower spectral distortion (MCD 154.6970) and better subjective sound quality (PESQ 1.1579) compared to MMS (MCD 162.4576, PESQ 1.0449). Despite lower intelligibility (STOI = 0.0444 vs. 0.2430 in MMS), KazakhTTS demonstrates high adaptability to the Kazakh language and is preferable for speech synthesis tasks within the framework of building a parallel corpus.