Speech-to-Text (STT)

You will see three labels for one capability: speech-to-text, ASR (automatic speech recognition) and, occasionally, voice recognition, which strictly means identifying who is speaking rather than what was said. Cloud providers tend to sell 'speech-to-text' APIs; papers and benchmarks say ASR; the underlying models and the metric, word error rate (WER), are shared. The reverse direction, text-to-speech (TTS), is the voice the agent speaks with.

Why the distinction matters less than the evaluation: a voice agent is a pipeline. Audio goes to STT, the text goes to the language model, the reply goes to TTS. If STT hears 'بكرة الساعة أربعة' as 'بكرة الساعة أربعة عشر', the best language model in the world books the wrong slot. Arabic makes this harder because most training data is Modern Standard Arabic while callers speak Najdi, Hijazi, Emirati or Egyptian, often mixing in English words, over a compressed phone line with background noise.

What to ask a vendor: What is your word error rate on my dialect, on phone audio, measured on real calls rather than a clean benchmark? Do you support streaming transcription so the agent can respond while the caller is still finishing (this drives latency and barge-in)? How are numbers, dates and Arabic-English code-switching handled? Nano AI evaluates STT engines per client on recordings from the client's own lines before choosing one, and re-checks the error rate as part of ongoing monitoring, because engines change and so do callers.