Speech-to-Text

Speech-to-Text is the process of converting spoken audio into written words using automatic speech recognition (ASR) models. A typical pipeline captures a waveform, applies a Mel spectrogram, and feeds the features into a neural network—often a Transformer like Whisper, Conformer, or wav2vec 2.0—that outputs token sequences. Post-processing fixes punctuation, casing, and speaker diarization, while language-model rescoring boosts accuracy on domain jargon. Modern APIs offer real-time streaming, word-level timestamps, and automatic translation, enabling captions, voice UIs, and meeting notes. Accuracy hinges on audio quality, accent diversity, and latency constraints; evaluation relies on word-error rate (WER) and real-time factor (RTF). Fine-tuning on recordings and vocabulary from a specific domain lowers WER, though the size of the gain depends on baseline accuracy, audio quality, and the amount of training data, and on-device models keep audio out of third-party processors, which removes some GDPR obligations around international transfers and sub-processors but does not by itself constitute GDPR compliance. By turning voice into searchable, analyzable text, Speech-to-Text powers chatbots, analytics dashboards, and Retrieval-Augmented Generation (RAG) pipelines.

Related terms

• What is Text Speech?
• What is Text to Speech used for?
• Amazon Titan Text
• Speech synthesizers use to determine context before outputting
• Split-and-Merge Retrieval

Related services: Agentic AI consulting.

← Back to the AI Glossary