Capabilities
Speech-to-TextSTT
Also known as: automatic speech recognition, ASR
AI that turns spoken audio into written text; the technology behind voice typing, captions and voice assistants.
Speech-to-text is AI that listens to talking and writes it down. It powers dictation on your phone, automatic subtitles on videos, meeting transcripts and the "listening" half of voice assistants. Modern versions are impressively accurate, even with accents and background noise, and can handle many languages, which is why voice is becoming a natural way to work with AI.
Speech-to-text (STT), also called automatic speech recognition (ASR), converts spoken audio into written text. It is the input side of voice interfaces (the output side is text-to-speech).
What it is used for
- Dictation and voice typing.
- Captions and subtitles.
- Meeting and call transcription.
- The first step in voice assistants, before an LLM processes the text.
Why it got so much better
Training on huge, varied audio datasets made models robust to accents, noise and many languages. Combined with multimodal AI, systems can now go straight from voice to a spoken answer.
Modern ASR uses end-to-end neural models, largely transformer-based encoder-decoders trained on large, weakly-supervised multilingual audio-text corpora (e.g. Whisper; Radford et al. 2022), replacing older HMM-GMM and hybrid pipelines. Training at scale yields robustness to accent, noise and domain, plus multilingual transcription and translation, without task-specific fine-tuning.
Design and evaluation concerns include word error rate (WER) and its limitations, streaming vs offline decoding (latency for real-time use), speaker diarisation, timestamps, and handling of code-switching and low-resource languages. STT increasingly feeds directly into LLM pipelines (voice agents) or is unified within multimodal models that ingest audio tokens, blurring the boundary between recognition and understanding. Failure modes, mistranscriptions on domain jargon, names and noisy audio, propagate downstream, so confidence handling and correction matter in production voice systems.