← all papers · overview

Transformer-Based Neural Network Approaches for Speech Recognition and Synthesis in the Sakha Language

Abstract

Recent breakthroughs in artificial intelligence and deep learning have fundamentally transformed the landscape of spoken language processing technologies. Automatic speech recognition (ASR) and text-to-speech (TTS) synthesis have emerged as essential components driving digital accessibility across diverse linguistic communities. The Sakha language, representing the northeastern branch of the Turkic language family, continues to face substantial technological barriers stemming from insufficient digital resources,  limited  annotated  corpora,  and  the  absence  of  production-ready  speech processing systems. This comprehensive investigation examines the feasibility and effectiveness of adapting contemporary transformer-based neural architectures for bidirectional speech conversion tasks in Sakha. Our research encompasses detailed analysis  of  encoder-decoder  frameworks,  specifically  OpenAI’s  Whisper  large-v3 and  Meta’s  Wav2Vec2-BERT  for  voice-to-text  transformation,  alongside  Coqui’s XTTS-v2  system  for  text-to-voice  generation.  Particular  emphasis  is  placed  on addressing linguistic and technical obstacles inherent to Sakha, including its complex agglutinative  morphological  structure,  systematic  vowel  harmony  patterns,  and distinctive phonemic inventory featuring sounds absent from most Indo-European languages.  Experimental  evaluation  demonstrates  that  comprehensive  fine-tuning of  Whisper-large-v3  achieves  exceptional  recognition  accuracy  with  word  error rate (WER) of 8%, while the self-supervised Wav2Vec2-BERT architecture attains 13%  WER  when  augmented  with  statistical  n-gram  language  modeling.  The neural synthesis system exhibits robust performance despite minimal training data availability, achieving average loss of 2.49 following extended training optimization and practical deployment via Telegram messaging bot. Additionally, ensemble meta-stacking combining both recognition architectures achieves 27% WER, demonstrating effective  complementarity  through  learned  hypothesis  arbitration.  These  findings validate transfer learning methodologies as viable pathways for developing speech technologies serving digitally underrepresented linguistic communities.

Related papers

Ranked by semantic similarity — how closely each paper's abstract matches this one (100% = near-identical topic).