Fiund / Speech & conversation data
ASR training data
Conversational speech for recognition models.
A clear review process
What to establish for your project.
Define the recognition task
Describe the speech the system will encounter: language, region, speaker variation, channel, recording environment and spontaneous versus read delivery. Separate training, fine-tuning and evaluation requirements so the proposed collection can be checked against the intended role.
Specify transcripts and alignment
State whether transcripts should preserve disfluencies, interrupted words, overlapping speech and non-speech events. Distinguish utterance-level timing from word-level alignment, and agree on the label conventions rather than treating a transcript file as a complete specification.
Review speaker and source groups
Identify repeat speakers, alternate episode edits and related source sessions. Define held-out groups for evaluation and the metadata needed to detect source leakage or near duplicates.
Start with a representative review
Include difficult examples in the proposed sample: accents, quiet turns, overlap, background sound and changing speakers. Discuss source documentation and permitted uses alongside technical acceptance before scoping a larger delivery.
Founder, Fiund
Your project contact
For research, engineering & procurement teams
A direct conversation.
A well-defined next step.
Speak directly with Jaeden Schafer about the recordings, documentation and preparation your speech-data project needs. Start with the model task and agree on a clear source-review and commercial scope.