Your reading path
Start with the right questions.
From principle to practice
Illustrative experiment: three nested subsets
Open the worked exampleFirst distinguish adaptation from training from scratch
Adapting an existing recognition model and learning a model from a new initialization are different experiments. A model’s previous training affects what the new examples need to contribute. A transcription service, a multilingual recognizer and a single-speaker synthesis project also use different targets, so their hour counts should not be compared as one market category.
Published corpus sizes describe particular research resources. For instance, LibriSpeech documents approximately 1,000 hours of read speech; its existence does not establish a minimum for another project. Start by naming the model, task and baseline evaluation result rather than importing an hours rule from an unrelated experiment.
Define the decision before buying a tranche
Write down a measurable failure: incorrect delivery dates in customer calls, missed technical terms in interviews, or poor recognition on a particular recording channel. Decide which evaluation slices represent that failure and how many independent speakers or sessions support them. Keep the evaluation set outside collection and model-selection decisions as far as practical.
Also define the operational consequence. A small change in average word error can coexist with unchanged errors on account numbers. A synthesis project may need pronunciation review and listener judgments rather than recognition WER. Choose a measurement that reflects the product and record where it remains an imperfect proxy.
Illustrative experiment: three nested subsets
The following plan is an experiment design, not reported performance or a recommended universal purchase size. Nested subsets mean the 5-hour set is included in the 10-hour set, which is included in the 20-hour set. The evaluation recordings remain fixed throughout.
- Run
- Baseline
- Data change
- No new task-specific training
- What stays controlled
- Evaluation IDs and scoring policy
- What to inspect
- Current error categories and slice counts
- Run
- A
- Data change
- 5 reviewed hours
- What stays controlled
- Starting checkpoint, recipe and stated compute budget
- What to inspect
- Whether the task can improve under this setup
- Run
- B
- Data change
- 10 reviewed hours including A
- What stays controlled
- Same evaluation and documented training changes
- What to inspect
- Marginal change and stability across runs
- Run
- C
- Data change
- 20 reviewed hours including B
- What stays controlled
- Same evaluation; record update counts
- What to inspect
- Whether the extra volume adds coverage or repeats the same sources
- Run
- Coverage check
- Data change
- A smaller alternative with new speakers or channels
- What stays controlled
- Comparable task definition and budget accounting
- What to inspect
- Whether diversity addresses the observed gap more effectively
Keep data, compute and composition from moving invisibly
If one experiment trains for more updates, uses cleaner transcripts and adds new domains at the same time, its result cannot be attributed to hours alone. Record each change and either hold it fixed or describe the experiment as a combined intervention. Equal epochs and equal optimizer steps control different things when dataset sizes change.
Report the actual composition: accepted speech duration, speaker/session counts, recording conditions and label review level. Twenty hours from one recurring host can offer fewer independent conditions than a shorter collection from new sources. Conversely, broad diversity can be unhelpful if none of it resembles the target task. The evaluation question decides which comparison matters.
Read the curve as a diagnostic, not a guarantee
Suppose a hypothetical score improves between A and B but changes little between B and C. That observation could indicate insufficiently novel data, inconsistent labels, inadequate optimization, measurement noise or a model limit. It does not prove a universal law of diminishing returns or identify one correct corpus size.
Repeat key runs when feasible and report variability instead of presenting one lucky result as a stable gain. Inspect error examples from each important slice and keep their denominators visible. A reduction measured on a very small subgroup may warrant another review sample before a purchasing decision.
Count usable speech and review work separately
State whether purchased hours mean wall-clock files, speech-bearing regions, distinct source duration or summed speaker time. Overlapping speakers and repeated derivatives can change those totals. Keep rejected files and excluded music or silence visible in a reconciliation report, rather than silently counting them as model-ready material.
Annotation effort is another axis. Correcting systematic transcript omissions may be more valuable than adding raw hours with the same defect. Ask for a pilot that tests the annotation policy as well as the acoustic coverage, and make quote units explicit before comparing providers.
Choose a stopping or expansion rule
Before running the experiment, state the conditions for the next decision. Examples include an agreed error reduction on a business-critical slice, stable performance across independently grouped evaluation sources, or a cost ceiling for accepted examples. These are project criteria, not thresholds supplied by this article.
If the target behavior remains weak, identify the next intervention with the strongest evidence: targeted recruitment, new environments, adjudicated labels, a model adjustment or additional analysis. AIPodcast can discuss sourcing against those gaps. There is no fixed hour count, delivery promise or guaranteed accuracy improvement implied by an inquiry.
Questions buyers ask
Sources & research notes
Public documentation reviewed 2026-11-13. Provider descriptions and published terms are attributed below. Published facts are attributed to their sources. Worked files are labeled as illustrative or as an inspected editorial source; neither implies commercial dataset clearance.
- Transformers automatic speech recognition task guideHugging Face · Reviewed 2026-11-13
One implementation recipe, not universal corpus specifications, data-hour requirements, or annotation-accuracy thresholds.
- LibriSpeech ASR corpus, SLR12OpenSLR / corpus authors · Reviewed 2026-11-13
Read-speech benchmark; download availability does not answer every project-specific privacy or voice-use question.
