Year: 2025
Type: article
Abstract: In the field of speech recognition, end-to-end models are gradually replacing traditional and hybrid approaches. Their main principle is autoregressive decoding, where the output sequence is formed from left to right. However, it has not yet been proven that this method provides the best results in converting speech to text. Moreover, end-toend models rely solely on the previous context, which complicates the processing of unclear or distorted sounds. In this regard, the insertion method was proposed, which does not use autoregressive decoding and generates output data in an arbitrary order. This paper examines a Kazakh speech recognition model trained using the insertion method and Connectionist Temporal Classification (CTC). The experiments conducted showed that this method improves recognition accuracy. Unlike autoregressive models, the Insertion method provides greater flexibility in processing sequences, as it does not require a strict order for generating output data. This reduces decoding delays and makes the model more robust to poorly pronounced words. Furthermore, combining the Insertion method with CTC improves the alignment of audio data and text transcription. This is especially important for agglutinative languages such as Kazakh. According to the experimental results, the recognition accuracy of the proposed model reached 10.2%, making it competitive today.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7117157127