Your reading path
Start with the right questions.
From principle to practice
Worked prompt: written text is not the spoken target
Open the worked exampleStart with the language variety and text pipeline
Name the language variety, pronunciation inventory, text normalizer and phonemizer versions before counting phones. The same written token can be expanded or pronounced differently across contexts and varieties. A count without those choices is difficult to reproduce or use for selecting new prompts.
The Coqui dataset-formatting guide discusses paired clips/text and phonemic variety. It does not establish one universal frequency target for every sound or a number of hours that guarantees a polished voice. Treat coverage as one diagnostic alongside recording consistency, pronunciation and the intended synthesis behavior.
Worked prompt: written text is not the spoken target
Consider the invented prompt “Dr. Lee paid $16.” The written form contains an abbreviation, a name and a currency amount. A US-English reading could be “Doctor Lee paid sixteen dollars.” Another context might require a different title expansion or spoken currency form. Store the intended expansion instead of expecting an annotation vendor or normalizer to guess it.
- Layer
- Written prompt
- Illustrative value
- Dr. Lee paid $16.
- Review question
- Which expansions and name pronunciation are intended?
- Layer
- Spoken target
- Illustrative value
- Doctor Lee paid sixteen dollars.
- Review question
- Does the speaker actually say this form?
- Layer
- Phone sequence
- Illustrative value
- Produced by the declared inventory and tool
- Review question
- Which neighboring-phone contexts occur or remain absent?
- Layer
- Recording label
- Illustrative value
- Reviewed wording and pronunciation notes
- Review question
- Was there a substitution, hesitation or reread?
Count contexts as well as individual phones
A marginal phone count says how often a phone occurs; it does not show its neighbors, word position or stress context. Repeating one familiar word can inflate a rare phone’s count while leaving other contexts uncovered. Keep context-level counts when they are relevant to the selected representation, and distinguish inventory coverage from natural frequency distribution.
For example, a prompt pack dominated by dates and prices may cover many digits while missing ordinary questions or phrase-final contrasts. Create a gap list tied to the product’s vocabulary and sentence functions. Do not force every theoretically possible phone sequence into the corpus: many combinations are absent from the language or irrelevant to the task.
Add prompts for a named gap, then audit the recording
Suppose a pilot reveals weak coverage of a selected phone near a word boundary. Write several natural prompts that vary the neighboring sound, word position and sentence function instead of repeating one tongue-twister. Record why each prompt was added so later reviewers can tell deliberate coverage work from arbitrary script growth.
After recording, compare the spoken text with the intended prompt. A speaker may normalize a number differently, omit a word, substitute a name or restart a phrase. Update the reviewed target and the realized-coverage counts accordingly. Planned coverage and accepted recorded coverage should be reported separately.
Prosody is another axis, not a phone-count correction
A corpus can contain the chosen phones while offering little variation in questions, emphasis, hesitation, pace or expressive style. Decide which functions the product needs and annotate them only with a clear convention. Treat subjective style labels as reviewed judgments rather than physical measurements.
Keep voice identity, session conditions and performance instructions visible. A stressed delivery can change timing and pronunciation, while a microphone change can create an acoustic difference unrelated to intended expression. A pilot matrix should help separate those factors so additional prompts address the actual gap.
Use alignment to inspect, not certify, coverage
Forced alignment can help locate supplied phone or word sequences, but an estimated boundary does not independently verify the pronunciation. Review out-of-vocabulary entries, compressed intervals, mismatched words and difficult contexts against the audio. The MFA alignment workflow explains the role of pronunciation and acoustic resources.
Use reviewed examples to distinguish a dictionary mismatch from a recorded mispronunciation or a timing failure. Preserve the correction reason and keep lexicon changes versioned. Otherwise the same recordings can appear to gain or lose coverage because the analysis tool changed, rather than because any new speech was collected.
Reserve a meaningful evaluation script
Avoid evaluating only the prompts used for coverage selection. Include held-out words, sentence structures and supported product cases, with a stated pronunciation reference. A few unseen sentences cannot establish general language quality, but they can expose whether a corpus improvement addresses the behavior it was designed to change.
Report the evaluation method and its limits. Listener judgments, pronunciation errors and task-specific measures answer different questions. Compare candidates under the same text-processing pipeline, and do not present a coverage percentage as a guaranteed synthesis-quality score.
What a TTS sourcing brief should carry
Include the intended language variety, script domain, text normalization examples, pronunciation inventory, desired sentence functions, session conditions and review protocol. Ask whether source voice permissions cover the proposed synthesis use; a public recording or an ASR transcript does not answer that question.
Availability, speaker participation, permitted uses, annotation work and delivery format need source-specific agreement. AIPodcast does not promise a standardized voice corpus or universal training-hour threshold through this guide. A well-defined prompt audit makes the project review more concrete without substituting for actual recordings or evaluation.
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.
- Formatting your TTS dataset, TTS 0.22.0Coqui TTS project · Reviewed 2026-11-13
A project-specific formatter guide; no universal phoneme frequency target, hours-per-voice rule or commercial voice permission.
- MFA 3.4.3 alignment workflowMontreal Forced Aligner · Reviewed 2026-11-13
Produced boundaries remain estimates; no universal timing-error tolerance is established.
