Speech Recognition

Speech recognition, or ASR, converts spoken language into text for voice interfaces, transcripts, and accessibility.

Speech varies by accent, speed, emotion, and microphone quality. Words smear together without clean edges. Noise sits on top. Older ASR stacked an acoustic model, a language model, and a pronunciation dictionary, each trained on a different objective. 0, and Conformer map waveforms to text with transformers, which simplifies the pipeline and raises accuracy.

Whisper, from OpenAI, is strong on multilingual speech. It trained on 680,000 hours of web audio covering 97 languages. Audio becomes a spectrogram, transformer encoders process it, then a decoder writes text.

Streaming models emit text as speech arrives. Speaker diarization labels who spoke when. Uses include voice assistants, meeting notes, video captions, call-center analytics, voice typing, and tools for deaf and hard-of-hearing users. Word Error Rate (WER) compares recognized words to a ground-truth transcript.

An acoustic model used to map frames to phonemes, a pronunciation dictionary mapped phonemes to words, and a language model scored word sequences. 0, and Conformer replace that split with one sequence model. Whisper’s 680,000 hours and 97 languages are why it holds up across accents and languages. Spectrograms are the usual input grid.

Streaming ASR emits partial transcripts before the speaker stops. Diarization adds speaker labels on top of the word stream. WER counts substitutions, deletions, and insertions against a reference. Deep Speech (2014) showed end-to-end speech-to-text with neural nets. Whisper (2022) later made multilingual ASR widely available.