Skripsi
Tdnn-f + hmm vs transformer for English speech recognition on edge device / Azarya Aditya Krisna Moeljono
Abstrak
Implementing Automatic Speech Recognition (ASR) on edge devices faces the engineering challenge of balancing transcription accuracy with computational efficiency. This study empirically compares the performance between hybrid ASR architectures namely Time-Delay Neural Network-Factorized Hidden Markov Model (TDNN-F HMM) implemented through Vosk with a modern end-to-end architecture namely Transformer implemented using the tiny variant Whisper model. The test was conducted directly on the Raspberry Pi 4 Model B edge device using a subset of data from the LibriSpeech corpus by evaluating two main metrics Word Error Rate (WER) for accuracy and execution time for efficiency. The experimental results showed that the Transformer (Whisper) architecture consistently achieved superior accuracy with an average WER score of 0.096 compared to 0.108 for Vosk which represented a decrease in the relative error rate of 11.1%. However in terms of efficiency the TDNN-F HMM (Vosk) architecture proved to be significantly faster with an average execution time of 4 043 seconds almost 80% faster than the Whisper which clocked 7 290 seconds. The study concluded that there is a clear trade-off between the two architectures Whisper offers higher accuracy while Vosk provides much lower latency. These findings provide essential evidence-based guidance for developers in selecting the optimal ASR architecture according to specific use case priorities whether for applications that demand high precision or those that require real-time responsiveness.