The short definition
An ASR dataset is a collection of audio recordings of speech paired with transcripts of what was said. Automatic Speech Recognition converts spoken language into text, so training requires examples of both sides of that mapping — the sound and the words.
In practice a usable dataset is more than audio plus text. It carries segmentation, timestamps, speaker information and recording metadata, all following consistent conventions.
What is inside a well-built dataset
- ▸Audio at a defined sample rate, bit depth and channel layout
- ▸Transcripts following documented conventions for numerals, disfluencies and foreign words
- ▸Utterance-level segmentation with timestamps
- ▸Speaker metadata: gender, age band, region, accent
- ▸Recording metadata: device, environment, noise level, distance
- ▸Manifests and documentation so the data loads reproducibly
Why coverage decides accuracy
Recognition quality collapses on speech that looks unlike training data. If a deployment will meet regional accents, noisy rooms, phone channels or fast spontaneous speech, those conditions need representation in the training set.
That is why teams commission collection instead of relying on found audio: coverage can be specified, quotas enforced, and evaluation slices designed in advance.
How the data is collected
Collection runs from a written specification: speaker profile, prompt or scenario design, recording conditions, transcription conventions and delivery format. Speakers are recruited and screened, a pilot batch confirms conventions, then the full set is recorded, transcribed, annotated and quality-controlled.
For Thai specifically, tone, the lack of word spacing and strong regional variation make specification and transcription conventions especially important.