Build a custom voice AI dataset: from a product error to an accepted release Translate a speech-model failure into a task contract, capture matrix, reference policy, grouped evaluation and pilot acceptance decision. A custom voice dataset should resolve a defined model or product gap. Start with the behavior that fails, specify what a correct example must contain, and test a small end-to-end pilot before scaling collection. The deliverable is a reviewable set of recordings, references and source records under an agreed use scope—not a nominal number of hours or a particular microphone specification. Turn the product failure into a data question Consider an invented delivery-support recognizer that misses corrected dates and parcel identifiers. “We need better English audio” is too broad. A more useful data question is whether the model handles a spoken correction under the intended device conditions and preserves the final identifier accurately. Decide whether the target is recognition of the entire utterance, extraction of the corrected value, or management of a live conversation. Those tasks may share recordings but require different labels. Write the input and expected output contract before commissioning the material. Create an example that exposes the distinction In the invented utterance “Send it Friday—sorry, Thursday, parcel A nine two,” the transcript represents all the spoken words. A final-date field represents Thursday. A dialogue label might describe a self-correction. These targets are related but cannot be substituted for one another. { "source_session": "pilot-014", "task": "recognition_and_final_value_review", "transcript": "Send it Friday—sorry, Thursday, parcel A nine two.", "final_date_spoken": "Thursday", "parcel_spoken": "A nine two", "correction_present": true, "reference_status": "illustrative_only" } Build a matrix around plausible failure causes Choose relevant combinations of channel, speaking style, language variety, environment and task content. Do not create an enormous Cartesian product that cannot be collected or reviewed. Identify the combinations the product actually expects and label untested intersections as gaps. For the delivery example, compare clean and deployment-like devices while keeping some utterance patterns comparable. Add new speakers and sessions rather than asking one speaker to read every combination. Keep task familiarity and script instructions visible so the project does not confuse rehearsed prompts with spontaneous repairs. Write the reference policy before labeling Specify verbatim versus cleaned text, spoken versus normalized numbers, overlap conventions, uncertain spans and the output format for a correction. Show both accepted and rejected examples. A team cannot resolve contradictory annotation instructions by collecting more audio. The [Hugging Face ASR guide](https://huggingface.co/docs/transformers/en/tasks/asr) and [Coqui formatting guide](https://docs.coqui.ai/en/dev/formatting_your_dataset.html) illustrate different preparation workflows. Use the selected model’s actual requirements; neither guide creates one universal WAV rate, transcript schema or minimum duration for every voice task. Design the evaluation boundary before the pilot expands Reserve source groups that should remain independent: complete sessions and known recurring speakers, and sometimes device setups or source organizations. A fixed five-percent speaker holdout is not automatically appropriate for every corpus. Choose the boundary and size according to the intended comparison and uncertainty. Track the grouping evidence. Unknown speakers cannot become confidently disjoint simply because files have different names. Inspect near-duplicate edits and derivatives and keep them with the relevant parent group. Avoid selecting evaluation examples because they make an initial model result look favorable. Agree go, repair and stop criteria A pilot should test the collection process, reference consistency and relevance to the target error. Define which defects block acceptance, which can be repaired and which are documented task conditions. Record stage counts so the buyer knows how much material remains after review. Pilot outcome Evidence Possible next action Target examples exist and are reviewable Sources, references and grouping resolve the intended question Expand only the identified coverage gaps Capture works but text policy is inconsistent Adjudicated examples expose conflicting label rules Repair the policy and re-review the pilot The source cannot reproduce the target behavior Reviewed samples differ from the deployment task Change the source or collection design Use or handling scope remains unresolved Specific evidence or agreement is still missing Resolve that dependency before release Package the result so another team can inspect it Provide source and segment IDs, files, accepted references, source clock/channel definitions, preparation records, split manifest and exceptions. Add a data dictionary explaining unknown, absent and excluded values. Keep the raw source distinct from resampled, denoised or segmented derivatives. A corpus validation pass, such as the [MFA validation workflow](https://montreal-forced-aligner.readthedocs.io/en/v3.4.3/user_guide/workflows/validate.html), can expose structural issues. It does not prove that a label, identity or proposed use is correct. The final release review should reconcile the actual artifacts with the agreed task and source scope. Use the first model result to choose the next collection Compare the baseline and pilot-trained result on the fixed evaluation set under documented controls. Look at corrected dates and identifiers as well as broad recognition errors. If performance remains weak, investigate references, source mismatch and optimization before assuming the answer is another block of hours. AIPodcast can discuss a project specification, candidate sources and preparation. Availability, participant involvement, intended use, review work and pricing require project-specific confirmation. No fixed capture equipment, delivery timeline or training-hours guarantee is established by this example. Should every custom project use studio microphones? No. Capture conditions should serve the task. A deployment-like channel may be essential for one recognizer, while a synthesis workflow may prioritize a different recording setup. Can one recording support several targets? Yes, if each target is defined and reviewed separately. A transcript, a final extracted value and a dialogue label answer different questions. What should a pilot prove before scaling? That the source produces relevant examples, the references are consistent, grouping can be documented and the remaining work and use scope are understood. Research method: Fiund research desk; public sources, explicit worked-example labels and project-specific availability review. Editorial contact: hello@fiund.com. Sources Transformers automatic speech recognition task guide: https://huggingface.co/docs/transformers/en/tasks/asr Formatting your TTS dataset, TTS 0.22.0: https://docs.coqui.ai/en/dev/formatting_your_dataset.html MFA corpus validation and first steps: https://montreal-forced-aligner.readthedocs.io/en/v3.3.4/first_steps/