Home › ASR training data
ASR training data that sounds like your users, not your booth
Speech recognition fails on the people it never heard. Accent, age, disfluency and background are not edge cases - they are most of your traffic. We collect the speech that actually turns up.
What we collect
- Spontaneous conversational speech - unscripted, two-party and multi-party, with the overlap and repair behaviour real dialogue contains.
- Prompted but natural speech - scenario-driven rather than script-driven, so the words are the speaker’s own.
- Accent and dialect coverage across the languages and accents we recruit, including second-language and code-switched speech.
- Demographic breadth - age bands, regional cohorts and speaker groups routinely thin in general corpora.
- Condition variation - device, distance and acoustic environment matched to where your system is deployed.
Directed, not just captured
Getting natural speech out of someone who knows they are being recorded is a skill, not a setting. Left alone, contributors perform: they slow down, over-articulate, and produce exactly the clean read that is useless for recognition training. Sessions are directed to get past that, which is where a background in broadcast production earns its place - our parent studio group has been directing unscripted audio for years before anyone called it training data.
Evaluation sets, held properly
If your evaluation set was recruited from the same pool as your training set, your word error rate is telling you what you want to hear. We can recruit and record held-out evaluation cohorts with deliberately different speaker profiles, so the number you report means something.
Consent, on the record
Conversational collection raises consent questions that scripted work does not - what was said, who else was present, what happens to it. Every contributor gives written, informed consent naming your organisation, the purpose, the territory and the licence term, and the provenance trail is built for Article 10 data governance review.
ASR collection questions
What audio does ASR training actually need?
The opposite of TTS. Real speech with real problems: disfluency, overlap, false starts, background, varied distance from the microphone, and the accents your users have rather than the ones the model was built on. Pristine studio reads produce models that fail the moment they meet a customer.
Why not just use scripted prompts?
Because people do not read prompts to your product. Read speech has different prosody, different pacing and almost none of the repair behaviour - “no, sorry, I meant” - that dominates real interaction. A model trained on read speech looks excellent in evaluation and fails in deployment.
Can you collect in noisy or in-car conditions?
Yes. Acoustic conditions are part of the specification: room type, distance, device, background profile. Controlled realism is the goal - conditions that match deployment, captured consistently enough to be usable.
Do you provide transcription and annotation?
Yes, to your schema - verbatim or clean-read transcription, speaker labelling, timestamps, disfluency and non-speech event tagging, and whatever else your pipeline expects. Annotation conventions are agreed before collection, not negotiated after.
Send us the failure cases
Tell us which speakers your model gets wrong. That is the collection brief.