Speech To Text: Definition and Examples

Speech To Text (literally "from speech to text") refers to all technologies capable of transforming an audio signal containing human speech into a textual transcription. This technology relies on deep learning models that analyze sound waves, identify phonemes, and then reconstruct the corresponding words and sentences.

Modern STT systems mainly use neural network architectures such as Transformers, which have significantly improved transcription accuracy. Models like OpenAI's Whisper or Google Speech-to-Text services can transcribe in dozens of languages, handle accents, background noise, and even distinguish multiple speakers (diarization).

In the context of prompt engineering and generative AI, Speech To Text often constitutes the first step of a larger pipeline. For example, a voice assistant first uses STT to understand the user's request, then a language model (LLM) to generate a response, and finally a Text To Speech system to vocalize it. The quality of the initial transcription directly impacts the relevance of the generated response.

Applications of STT are ubiquitous: automatic video subtitling, voice dictation, meeting transcription, accessibility for the hearing impaired, voice control of connected devices, or conversation analysis in call centers. The rapid evolution of these technologies makes automatic transcription increasingly reliable, approaching or even exceeding human accuracy in some contexts.