Skip to content

Buying training data

How much audio do you need for a TTS model?

Separate pretraining, fine-tuning and reference audio before choosing an hour count. Published corpus examples, an original sample brief and an evaluation plan.

For
AI data buyers
Reading time
6 min
Updated

Key takeaways

  • Separate corpus size, adaptation material and prompt length.
  • Use published dataset quantities as attributed context.
  • Specify source, transcript and review status per segment.
  • Evaluate the target behavior on held-out material before increasing volume.
  • Fiund’s standard terms prohibit identifiable voice or likeness replication.

There is no universal number of hours that makes a TTS project production-ready. The base model, target languages, speaker and acoustic coverage, transcript quality, evaluation criteria and permitted use all affect the collection you need.

First distinguish the complete training corpus from the smaller material used for adaptation or a reference prompt. A paper describing a short prompt may rely on a model pretrained on a much larger corpus. Those quantities answer different questions and should not be used interchangeably in a recording budget.

Published research quantities are context, not a recording prescription
ReferencePublished source scaleHow to read the comparison
LJ SpeechApproximately 24 hours of a single speaker, with matching transcripts.A concrete corpus example. Its size is not a universal threshold for another model or use.
LibriTTS585 hours from 2,456 speakers in the original paper.Corpus diversity and total duration differ from a single-speaker adaptation set.
VALL-E paper60,000 hours of pretraining speech and a three-second enrolled prompt.The prompt length does not describe the full training-data requirement.

Define the model job before estimating duration

State whether the work is training a model, adapting an existing one, evaluating synthesis or testing another speech task. Name the base model and version, target language and conditions, expected outputs and acceptance criteria. A duration estimate without those assumptions is difficult to evaluate or price.

If a published model reports a result on a particular corpus, preserve that context. Different preprocessing, splits, speaker populations or acoustic conditions can change the task. A corpus example is a starting point for investigation, rather than a promise that matching its hour count reproduces the reported performance.

Specify source and transcript relationships

A useful brief identifies original audio, channel structure, capture conditions, language, speaker references and matching text. Define how numbers, abbreviations, hesitations, punctuation and non-speech events are represented. Keep verbatim and normalized text distinct when both are used.

Document segment boundaries, missing or uncertain speech, transcript review and the relationship between source and processed files. A model-generated transcript should remain labeled as generated unless the defined human review has actually occurred. Track file versions so an evaluator can reproduce the package.

An original sample brief

Consider a fictional speech-evaluation project, SAMPLE-1. It asks for original audio and a checked transcript, with one shared segment identifier, a documented language and a record of permitted evaluation use. The supplied preview has readable audio and automatic text, but no transcript-review record. Mark the source audio and generated transcript as present and human review as unknown. Do not label the whole sample verified.

Add a reviewed comparison sample with a different session identifier and the same annotation convention. Record the differences in source conditions and review status. This exercise concerns documentation and evaluation readiness. It is not a paid order, a tested TTS outcome or a claim of available Fiund inventory.

Evaluate coverage before expanding the collection

Build a held-out evaluation plan around the behavior the project actually needs. Keep related recording copies and adjacent segments grouped appropriately so near-duplicate material does not make a test appear easier. Record the split method and any speaker or session separation required by the objective.

Compare performance, limitations and failure categories as relevant data coverage grows. Inspect pronunciation, naturalness or other task-appropriate outcomes with an agreed review method. A transcript error rate can be useful for a defined recognition subtask, but it is not a complete measure of synthesis quality. Do not replace the target evaluation with an easier metric simply because a tool reports it.

Keep podcast conversations and purpose-recorded speech distinct

A podcast can contain overlap, variable microphone conditions, music and edits. It can be useful for natural-conversation tasks while being a poor match for a different recording specification. Ask which properties are required before assuming that an existing show replaces a purpose-recorded corpus.

For an archive owner, inventory original tracks and available transcripts and connect them to permission records. Mark unknowns and excluded ranges. Original channel separation, estimated speaker turns and later source separation are different forms of evidence and should remain labeled.

Confirm permitted use and Fiund’s scope

Review the actual licence and participant permissions for the proposed source and task. Training, evaluation, redistribution and identity replication are separate questions. A public download or ordinary recording release does not settle every proposed use. Source-specific restrictions still need review.

Fiund’s standard contributor agreement prohibits a synthetic replica, clone or voice or likeness model intended to imitate an identifiable person, as well as use of identity to imply endorsement. A general speech-data request to Fiund must fit those restrictions. This guide explains data-planning distinctions and does not offer a voice-cloning service.

Turn the estimate into a comparable quote

Separate source licensing or collection from transcription, normalization, alignment, annotation, review and delivery. State whether the payable quantity is recorded duration, accepted segments or another agreed unit. Define the evidence and procedure for acceptance, rejection and rework before preparing a larger batch.

Request a representative sample and its manifest first. Record which requirements are supported, unsupported or unknown. The next quantity decision should follow the task evaluation, coverage gaps and review capacity, with the assumptions preserved for both buyer and supplier.

Fields for a speech-data sample brief

  • Model task and version, target language and intended output.
  • Required source conditions, channels and segment identifiers.
  • Transcript conventions, normalization and actual review status.
  • Coverage, split policy and acceptance examples.
  • Permitted uses, source authority and restrictions.
  • Included preparation, payable unit and delivery records.

Sources and further reading

Frequently asked questions

Is 24 hours enough to train a TTS model?

LJ Speech is an approximately 24-hour corpus, but its size is not a general production threshold. The model, source conditions, coverage, preprocessing and evaluation objective determine whether a collection fits.

Does a three-second reference mean training needs only three seconds?

No. The cited VALL-E paper separates a short enrolled prompt from a much larger pretraining corpus. Prompt length and full training-data requirements are different quantities.

Does Fiund license recordings for voice cloning?

Standard Fiund terms prohibit a synthetic replica, clone or voice or likeness model intended to imitate an identifiable person. A speech-data project must fit the actual permitted uses and restrictions.

Have a defined speech-data task?

Describe the intended use and source requirements →

Related resources

← Buying training data