Automatic speech recognition (ASR)
Also known as：ASR · speech-to-text · STT
Automatic speech recognition (ASR) is the technology that converts a speech signal into text. In live translation it is the first stage of the chain, and the text it produces sets the ceiling for everything that follows.
Streaming versus batch recognition
Batch recognition takes a complete audio segment and returns a result. Accuracy is usually higher, but nothing happens until the speaker stops, which rules it out for live use.
Streaming recognition emits output while audio is still arriving, producing provisional text during speech and revising it as more context appears. Live translation requires it.
Streaming output comes in two kinds:
- Partial results, which may still be overturned, useful for driving a scrolling live caption;
- Final results, which the system considers settled and ready to translate.
The distinction matters because translation can only act on final results, and waiting for a final result is itself part of your latency budget.
How it is measured
Character error rate (CER) is standard for Chinese, word error rate (WER) for English. Both are simply errors divided by total units.
Published accuracy figures are almost always measured on clean audio with standard pronunciation. Real studio numbers are typically much worse. The only figure worth acting on is the one you measure using your own broadcast audio.
What makes live streaming hard
Background music, elision at high speaking rates, out-of-vocabulary brand names and model numbers, and hosts verbally correcting themselves. None of these dominate in a meeting; all four appear at once in live selling.