KMITL

Permanent URI for this communityhttps://dspace.kmitl.ac.th/handle/kmitl/1

Browse

Settings

Subject: search.filters.subject.Automatic Speech Recognition

Search Results

Now showing 1 - 3 of 3
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    ASR Error Correction in Low-Resource Burmese with Alignment-Enhanced Transformers Using Phonetic Features
    (2025-01-01)
    Lin, Ye Bhone
    ;
    Aung, Thura
    ;
    Thu, Ye Kyaw
    ;
    Oo, Thazin Myint
    This paper investigates sequence-to-sequence Transformer models for automatic speech recognition (ASR) error correction in low-resource Burmese, focusing on different feature integration strategies including IPA and alignment information. To our knowledge, this is the first study addressing ASR error correction specifically for Burmese. We evaluate five ASR backbones and show that our ASR Error Correction (AEC) approaches consistently improve word- and character-level accuracy over baseline outputs. The proposed AEC model, combining IPA and alignment features, reduced the average WER of ASR models from 51.56 to 39.82 before augmentation (and 51.56 to 43.59 after augmentation) and improving chrF++ scores from 0.5864 to 0.627, demonstrating consistent gains over the baseline ASR outputs without AEC. Our results highlight the robustness of AEC and the importance of feature design for improving ASR outputs in low-resource settings.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Improving OpenAI's Whisper Model for Transcribing Homophones in Legal News
    (2024-01-01)
    Siriket, Lattapon
    ;
    Jitkajornwanich, Kulsawasd
    ;
    Jaiyen, Saichon
    ;
    Intakosum, Sarun
    The 'Whisper' model provides a tool for those who require transcription of human voice. It equips with opensource features and diverse functionalities. The model is capable of effectively deciphering messages in multiple languages, including support for the Thai language. This paper focuses on improving the transcription process of Thai homophones using the Whisper model in reducing the word error rate (WER). We focus on words in the legal news category and identify factors that lead to Whisper's incorrect sound predictions. We examined homophones using snippets of legal news video clips and compiled them into a homophone dictionary. We compare words extracted from the Whisper model by determining the word error rate and spelling of words. Based on the initial results obtained from the original Whisper model and the created homophone dictionary, 48 % of the words were incorrectly transcribed out of a total of 94 words. Then, we propose a methodology by which the performance of the Whisper is improved. That way, the automatic speech recognition of Thai language using the Whisper model can fully be utilized and used in other applications.
  • Loading...
    Thumbnail Image
    Some of the metrics are blocked by your 
    consent settings
    Item type:Publication,
    Comprehensive Benchmarking and Analysis of Open Pretrained Thai Speech Recognition Models
    (2024-01-01)
    Tipakasorn, Pattara
    ;
    Chatthong, Oatsada
    ;
    Yonehana, Ren
    ;
    Thangthai, Kwanchiva
    This paper presents a comprehensive benchmarking and analysis of open pretrained Thai Automatic Speech Recognition (ASR) models, addressing a critical gap in low-resource language ASR development. Our benchmarking focuses on one foundation speech model, three Open Thai Speech Recognition Models, and Open speech APls. We evaluate these models, including our fine-tuned Whisper model, across diverse speech types and acoustic environments. The study reveals significant performance gaps between read and spontaneous speech, with models performing well in controlled settings but struggling in real-world scenarios. We introduce evaluation datasets for distant speech and noisy podcasts, exposing limitations in current models' robustness. Our fine-tuned Whisper model demonstrates performance across various Thai regional dialects, reflecting its targeted training on dialectal data, while the others demonstrate resilience in spontaneous speech scenarios. However, all models show substantial degradation in challenging acoustic conditions, indicating a need for more diverse training corpora to better capture real-world complexity, including spontaneous speech and far-field acoustic scenarios, to further enhance Thai ASR. This work provides valuable insights for improving ASR models in low-resource languages.
      3