Automatic speech recognition (ASR)
Automatic speech recognition (ASR) is the branch of AI that recognizes spoken language in audio and converts it into text. ASR models learn the link between sound patterns and words from large collections of recorded speech. In video, ASR is the engine behind transcripts, captions, voice search and spoken-word indexing, and its accuracy is often measured as word error rate (the share of words it gets wrong).
Why it matters
ASR quality decides whether captions are usable, whether search finds the right moment and whether moderation catches spoken abuse. Poor recognition creates rework, accessibility complaints and missed content. Language coverage also limits which markets a product can serve without manual work.
How it works
- Audio is converted into a spectrogram, a picture of how sound energy changes over time.
- An acoustic model maps short slices of that sound to speech sounds or words.
- A language model chooses the most likely sequence of words in context.
- Post-processing adds punctuation, capital letters, formatted numbers and timestamps.
What is a good word error rate for ASR?
There is no single target; lower is better, and the acceptable level depends on the use. Word error rate counts substituted, deleted and inserted words, divided by the number of words actually spoken. For example, 8 errors in a 200 word passage is a 4% word error rate. Published captions usually still need a human pass on names and numbers.
What is the difference between ASR and NLP?
ASR converts speech audio into text. NLP (natural language processing) works with text that already exists, for example to summarize, translate or extract names. In a video pipeline they run in sequence: ASR produces the transcript, then NLP tools analyze it. Errors from ASR carry into every NLP step that follows.
Where it's used
Creator platforms run ASR on uploads to create captions and moderate what is said. News teams transcribe interviews within minutes, and SaaS meeting and support tools make recordings searchable.
Where it usually goes wrong, and how FastPix handles it
ASR output is often treated as final, yet it feeds captions, summaries and search, so recognition errors and missing languages spread across the product.
Recorded videos get transcripts and captions with multiple language tracks from FastPix's in-video AI, and live streams can carry real-time captions, so a team can caption both its library and its events without adding a separate vendor.
By the numbers
- OpenAI released its Whisper speech recognition model on September 21, 2022, trained on 680,000 hours of multilingual and multitask supervised data from the web. (Introducing Whisper (OpenAI))
