STT
An AI technology that converts voice signals into computer-readable text data. Recently, deep learning models such as Whisper and Gemini offer high accuracy even in noisy environments, making it useful for automatic subtitling and meeting transcription.
Detailed explanation
Why it matters in tool selection
When selecting an STT tool, the core metric is Word Error Rate (WER). Beyond simple recognition rates, one must consider background noise cancellation capability, domain-specific terminology recognition performance, and latency. Especially in meeting situations with multiple speakers, the performance of 'speaker diarization'—which accurately distinguishes each speaker—determines the level of workflow automation.
Key considerations
- Is the recognition accuracy high for Korean-specific honorifics and homophones?
- Does it support speaker diarization?
- Is the model installable in a local environment (on-premise) for security?
- Is the cost per API call or per hour suitable for your budget?
Use cases
Services like CLOVA Note utilize STT to transcribe meetings with multiple participants into text, which is then summarized by AI to generate reports. In addition, YouTube's automatic subtitle generation service provides near-real-time accessibility to viewers worldwide.
Easily confused terms
TTS (Text-to-Speech)
In contrast to STT, this technology outputs text data as human voice (audio).
NLP (Natural Language Processing)
A technology for understanding and analyzing text, used to grasp the context of the text transcribed by STT.