What is ASR?

An ASR system needs a large volume of paired audio and text to work at all. That pairing is the most basic form a speech dataset takes: one recording, one verbatim transcript.

Three things worth asking before buying ASR training data: how many distinct speakers are in it (not how many hours), what the recording conditions were, and whether the transcript convention follows pronunciation or standard orthography.

Types of ASR systems

ASR systems divide first by when they transcribe. An offline system takes a finished recording and returns the whole transcript, which lets it use later audio to correct earlier decisions. A streaming system emits words while the speaker is still talking, working only from the audio so far. It is judged on speed as well as accuracy.

The second split is what the speaker is doing. Read speech follows a script: well-formed sentences, few false starts, and a transcript that can follow ordinary spelling. Spontaneous speech is unplanned, full of restarts and filled pauses, and the transcript has to decide whether those are written down or dropped. A model trained on one and deployed on the other loses accuracy.

The third is vocabulary. A general-purpose system wants breadth: many speakers, many topics, ordinary words. A domain system such as medical dictation, call-centre audio or industrial voice commands wants depth in a narrow vocabulary, with the awkward items spoken aloud in the recording. Drug names, part numbers and place names do not appear in general conversation, so general audio will not teach a model to hear them.

What to specify when buying

The first number to ask for is the speaker count, not the hour count. A hundred hours from twenty speakers teaches twenty voices; the same hundred hours spread across five hundred speakers teaches how voices vary. Ask how many speakers are in the set, how many hours each contributes, and whether the recordings came from one device or several.

Then the transcript convention: verbatim or cleaned, whether fillers and stutters are written down, how numbers are spelled, how punctuation is placed. For call audio, ask how the channels are laid out. A recording with one speaker on each channel is ready to use as it stands; the same call mixed into a single channel needs diarization first, and that work is quoted separately.

Vagueness costs money in three predictable places. Hours quoted without a speaker count hide a narrow speaker pool. Two batches transcribed under different conventions cannot be pooled into one training set. And a delivery with no held-out test set leaves you unable to check the accuracy you were quoted, because the only party who can measure it is the supplier.