Skip to content

Latest commit

 

History

History
94 lines (73 loc) · 4.42 KB

File metadata and controls

94 lines (73 loc) · 4.42 KB

Pronunciation and speech text

Pandrator keeps display text separate from the text sent to a speech provider. That distinction allows speech-specific respelling, abbreviation expansion, or optimization without silently changing subtitles, chapters, or source documents.

Three different edits

Mechanism What changes When to use it
Edit or Search/Replace The selected source, display, or speech revision The stored text itself is wrong
Pronunciation library Only the synthesized request payload A known written term needs a deterministic spoken form
Speech-text optimization A separate, reviewable speech revision Broader provider-specific preparation needs language judgment

Do not use pronunciation rules to conceal errors that should be corrected in the actual document or subtitle track.

Add a pronunciation

  1. Open Pronunciations and add the written form plus a lowercase structured respelling.
  2. Leave it as Needs review until you have listened to it.
  3. Approve it as Reviewed and choose global or one-session scope.
  4. Optionally restrict it to a language and TTS backend. und matches any language and * matches any backend.
  5. In the session's deterministic text settings, keep Apply reviewed pronunciation-library overrides enabled.

Only reviewed entries in the applicable scope are active. Matching is bounded, case-insensitive, longest-first, and non-overlapping. The display text remains unchanged; the replacement occurs when Pandrator prepares a synthesis request.

After changing an entry, regenerate affected audio. If you deliberately reuse an older saved speech plan, rebuild that plan too: Pandrator cannot infer and reverse respellings already stored in a historical speech revision.

Speech-text optimization

Optimization can prepare numbers, abbreviations, punctuation, pronunciation, or phrasing for a particular TTS service. Native execution uses the configured multi-provider LLM adapter; no particular provider such as Ollama is required. Passive MCP execution instead lets the model already running in the host claim and return speech-text batches without a Pandrator model call. Both create a distinct speech revision that must be reviewed.

Whole-document optimization and generation-time batch optimization are two places to perform similar work. Avoid enabling both accidentally. If an LLM proposes a pronunciation-library entry, it remains inactive until you review and approve it.

Guarded mode is the normal constrained speech planner. Flexible mode allows more speech-only punctuation and phrasing freedom inside a protected template. The retired prompt-rewrite mode is no longer a user choice. For a small local model, set units per model request to 1; a capable model can process larger requests while Pandrator still validates every unit independently.

See the speech-optimization reference for standalone versus generation-time behavior, batching, settings, source formats, and the passive dispatch contract.

Mixed-language speech

Most local wrappers accept one language for one synthesis request. Tags such as [en]…[/en] are normally sent literally and are not a portable code-switching protocol.

When practical, isolate a foreign phrase as its own segment and use a per-segment language/voice override or Generate alternate take. The alternate take can select provider, model, voice, language, speech prompt, and RVC settings when the service supports them. A compatible multilingual voice may work across the selected languages; otherwise choose an appropriate voice for each segment.

Surgical regeneration of an inline word or phrase would require reliable span alignment, a compatible voice in both languages, silence and duration matching, and review confidence. Do not assume that Pandrator currently performs that splice automatically.

Review by listening

Text review cannot establish whether a speech-specific transformation sounds right. Generate short representative takes and listen for:

  • names, numbers, dates, acronyms, and abbreviations;
  • language transitions;
  • unwanted literal punctuation or tags;
  • stress, rhythm, and pauses;
  • regressions caused by a broader replacement; and
  • differences between original and RVC-converted takes.

For the surrounding generation flow, see your first audiobook and your first voiceover.