Speaker diversity in speech data: audit concentration and gaps Measure speaker counts, speaking-time concentration and evaluation coverage without inferring demographic attributes from voice. Includes a worked concentration example. Speaker diversity is not captured by a unique-speaker count alone. A corpus can contain many guests while a handful of recurring hosts provide most of its speech. Audit speaker hours, source relationships and the provenance of any coverage labels, then evaluate the conditions relevant to the product. Do not infer demographic identity from a voice to make a table look complete. This guide explains an operational concentration audit and the limits of the conclusions it supports. Measure together Unique speakers, speaking time and source families Metadata rule Record evidence and unknowns Evaluation unit Independent speakers and related sessions Avoid One aggregate score presented as universal coverage Separate variety from representativeness A dataset can vary across many voices without representing the people who will use a product. Conversely, a deliberately narrow collection may be appropriate for a limited internal task. State the intended population, interaction and channel before judging whether the source has enough relevant variation. Speaker count, language count, recording location and demographic coverage are different measures. A supplier may support some reliably while leaving others unknown. Ask how each field was obtained and whether its definition is consistent across sources. A global archive total cannot answer a narrowly defined deployment question. What disparity research supports [Koenecke and colleagues’ 2020 study](https://pmc.ncbi.nlm.nih.gov/articles/PMC7149386/) documented group differences in the ASR systems it evaluated. It is a reason to examine performance gaps and relevant source conditions, not evidence that every present-day model has the same gaps or that one particular recruitment ratio fixes them. Keep those boundaries in public claims. A data proposal should identify what variation it can document and how a pilot will assess the intended failure modes. It should not advertise diversity as a guaranteed accuracy improvement, an insurance policy or proof of fairness. An illustrative speaking-time concentration audit Assume an authored example with 120 hours of accepted speech: four recurring hosts contribute 20 hours each, and forty guests contribute one hour each. There are 44 distinct speaker IDs, but the four hosts supply 80 of 120 hours, or two-thirds of the speech. Measure Illustrative value What it reveals Distinct supported speaker IDs 44 Breadth of identities in the manifest Hours from the four hosts 80 / 120 = 66.7% A small group dominates speaking time Hours from forty guests 40 / 120 = 33.3% Many speakers contribute relatively little Equal-time reference 120 / 44 ≈ 2.73 hours per speaker A comparison point, not an optimal collection target Unknown or unresolved identity Report separately if present Prevents uncertain IDs from being treated as verified new speakers A defined concentration index can add context For this example, let each speaker’s share of total speech be p. The inverse concentration value is 1 divided by the sum of p squared. Four shares of 20/120 and forty shares of 1/120 give approximately 8.78. That means the observed time concentration has the same index value as about 8.78 equally represented speakers, despite the nominal count of 44. This is descriptive arithmetic, not an industry acceptance threshold or a measure of demographic fairness. It does not account for topic, accent, channel or task difficulty. Its value is that it exposes a distribution the raw speaker count hides. A project can choose other concentration measures if the method and denominator are stated. Use documented attributes without inventing identities [FLEURS’ dataset card](https://huggingface.co/datasets/google/fleurs) illustrates how language and other fields can be made explicit in a public corpus. For a proposed commercial source, inspect the actual label definitions and collection method rather than copy the benchmark’s categories into missing metadata. Self-reported information, administrative records and automatic inferences have different evidential status. Unknown values should remain visible. Speaker names generally are not needed to publish an aggregate coverage report. If identifying or sensitive information is proposed, assess its necessity, handling and applicable obligations with the responsible reviewer; [GDPR](https://eur-lex.europa.eu/eli/reg/2016/679) is one relevant framework where applicable. Protect evaluation from recurring voices and derivatives A clip-level random split may place the same host in training and testing. Alternate edits and repeated excerpts can create additional dependence. Define whether the evaluation is intended to test a familiar speaker, an unfamiliar speaker, an unfamiliar production or some combination, and group the material accordingly. Keep the actual grouping identifiers stable through segmentation and preparation. A hash of an output filename is not a speaker identity, and a different filename does not establish independent source material. Record unresolved identity relationships so evaluation reports can explain their limits. Report slice results with the denominators visible The [Transformers ASR guide](https://huggingface.co/docs/transformers/en/tasks/asr) uses WER as one recognition metric. An operational audit should report the reference convention and the number of speakers, segments and words behind each relevant slice, not just a decimal score. If a difficult slice contains two speakers and a few minutes of audio, say so. An apparent improvement may depend heavily on one session. Review errors and uncertainty at an appropriate independent unit rather than treating thousands of words from one host as thousands of independent people. Do not rank groups with unsupported identity labels. Turn the audit into a specific next action If a recurring host dominates the source, the answer is not automatically to discard that host or force equal hours. The source may still fit the target application. The audit should identify which failure or missing condition the next collection would address and what evidence a pilot needs to provide. Reconcile speaker IDs and record unresolved relationships. Report time concentration beside unique counts and source-family counts. Document label provenance and unknowns for the coverage fields that matter. Compare deployment-relevant groups while acknowledging channel and topic confounders. Keep evaluation groups independent according to the stated question. Request targeted additions only after identifying a concrete coverage or performance gap. Does a high speaker count prove representative coverage? No. It can conceal concentrated speaking time, repeated sources or unsupported attributes. Review the distribution and definitions against the intended users and conditions. Should every speaker contribute equal hours? Not necessarily. Equal hours are one reference distribution, not a universal optimum. Choose a sampling plan that fits the task and evaluate its consequences explicitly. Can demographic fields be inferred from the recording? Do not invent them for a coverage report. Use supported source metadata with its collection method and appropriate handling, and leave unavailable attributes unknown. Research method: Fiund research desk; public sources, explicit worked-example labels and project-specific availability review. Editorial contact: hello@fiund.com. Sources Racial disparities in automated speech recognition: https://pmc.ncbi.nlm.nih.gov/articles/PMC7149386/ FLEURS dataset card: https://huggingface.co/datasets/google/fleurs Regulation (EU) 2016/679, GDPR: https://eur-lex.europa.eu/eli/reg/2016/679 Transformers automatic speech recognition task guide: https://huggingface.co/docs/transformers/en/tasks/asr