AI BriefingAutomatic Speech Recognition (ASR)
Key point
ASR converts speech to text and has improved accuracy and real-time processing performance over the past 70 years.
Details
Automatic Speech Recognition (ASR) is a technology that listens to human speech and converts it into machine-readable text. It is widely used in everyday life and services, including voice assistants, message dictation, real-time captions, and IVR phone systems.
ASR is the technology that identifies what is spoken in audio, while STT (speech-to-text) is closer to the concept of implementing this as a feature or application. Speech recognition technology adds speaker identification to this, distinguishing who said what, which serves as the basis for speaker diarization.
The evolution of ASR is as follows:
- 1952: AUDREY, developed by Bell Laboratories, recognized digits from 1 to 9. The accuracy was approximately 90%, but it only worked when used by the inventor.
- 1960s–1970s: Research institutions expanded recognition vocabularies, and Raj Reddy's research laid the foundation for continuous speech recognition without pauses between words. DARPA research contributed to recognizing over 1,000 words in 1976 by utilizing Beam Search.
- 1980s–early 2010s: Combining Hidden Markov Models with Beam Search enabled systems that worked without training tailored to specific speakers. Dragon Systems released NaturallySpeaking Preferred, a commercial ASR capable of recognizing 100 words per minute, in 1997.
- Post-2010s: End-to-end neural networks that directly learn from speech and transcription data began to outperform traditional statistical pipelines. Currently, ASR is based on large-scale neural networks trained on millions of hours of audio, processing dozens of languages in real time.
A representative metric for ASR performance is the Word Error Rate (WER). In actual services, latency, speaker diarization quality, and contextual understanding ability also serve as important evaluation criteria alongside accuracy.
ElevenAPI provides production-grade ASR for developers and is introduced as having recorded the lowest Word Error Rate of 2.2% in Artificial Analysis's independent benchmark as of July 2026.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.