speech-to-text (stt / asr)

Speech-to-text (STT), also known as Automatic Speech Recognition (ASR), transcribes spoken audio into written text, converting acoustic signals into sequences of words. ASR is a foundational technology enabling voice interfaces, transcription services, and human-computer interaction through speech. ASR architectures have evolved through several paradigms: traditional pipeline (acoustic model mapping audio features to phonemes, pronunciation dictionary mapping phonemes to words, language model providing linguistic context — using Hidden Markov Models with GMMs or DNNs), hybrid models (combining deep neural networks for acoustic modeling with weighted finite-state transducers for decoding), end-to-end models (single neural networks mapping audio directly to text — CTC-based like DeepSpeech, attention-based encoder-decoder like Listen Attend and Spell, and RNN-Transducers like those used in streaming applications), and modern transformer-based models (Whisper by OpenAI — trained on 680K hours of multilingual supervised data achieving near-human accuracy across many languages, Conformer — combining convolution and self-attention, and wav2vec 2.0/HuBERT — self-supervised pre-training on unlabeled audio followed by fine-tuning). Key technical components include: feature extraction (converting raw audio to mel-frequency cepstral coefficients or mel spectrograms), language modeling (incorporating linguistic context to disambiguate acoustically similar words), beam search decoding (exploring multiple hypotheses simultaneously), and voice activity detection (identifying speech segments in audio). Challenges include: noisy environments (background music, multiple speakers, reverb), accented or dialectal speech, code-switching (speakers alternating between languages), domain-specific vocabulary (medical, legal, technical terms), real-time processing requirements for streaming applications, and speaker diarization integration (identifying who said what). Leading systems include Whisper, Google Speech-to-Text, Amazon Transcribe, Azure Speech Services, and AssemblyAI.