Automatic Speech Recognition (ASR) converts spoken language into written text by learning the relationship between audio signals and textual sequences. As shown in the figure, speech input is first transformed into a log-Mel spectrogram, which represents the acoustic features of the audio. A Transformer encoder then extracts high-level speech representations, while a Transformer decoder generates the corresponding text tokens through sequence-to-sequence learning. By training on large-scale multitask speech data, the model can handle English transcription, multilingual transcription, speech translation, and even no-speech detection. This makes modern ASR systems more robust, flexible, and capable of supporting diverse real-world speech understanding tasks.