Phoneticize — TTS pronunciation prep

A workflow for identifying pronunciation hazards in a TTS script, agreeing phonetic renderings with the user via audio previews, and emitting a phoneticized script ready for narration.

Read first

Governing principles

1. Preview the fragment, not the word. TTS prosody depends on surrounding context. A respelling that sounds right in isolation collapses inside a sentence. Render fragments throughout.
2. Respelling is the primary output, not SSML. Eleven v3 silently drops <phoneme> tags (see references/eleven-v3-notes.md). Inline respelling — Siobhán → shi-VAWN directly in the prose — is what actually changes the model's output. IPA goes in the table for archival precision; respelling goes in the script.
3. Mark uncertainty, never guess. A wrong respelling shipped with confidence is worse than an explicit ? the user resolves. When the pronunciation isn't obvious, ask.
4. Stable IDs across iterations. Each candidate gets a row ID (P01, P02, …) on first pass and keeps it for the lifetime of the table. The user references rows by ID; regenerate only the rows that changed.
5. Render once, accept once. Never re-render an accepted row — it wastes Higgsfield calls and the user will assume something broke when the audio differs subtly between takes.

Phase 1 — Scan

Combine the regex helper with semantic reading. Neither alone catches everything: regex misses ordinary-looking words with non-obvious pronunciation (Worcester, Featherstonehaugh), and an LLM scan alone will miss tokens deep in long scripts.

Run the helper

python scripts/extract_candidates.py path/to/script.txt --out candidates.json

The helper applies the patterns documented in references/detection-heuristics.md and emits a JSON list of deduplicated candidates with category guesses, positions, and context fragments.

Add and prune

Read the script. For every candidate the helper found:

• Confirm the category — the helper guesses from token shape and occasionally gets it wrong (a CamelCase brand name miscategorised as ART, etc.)
• Drop false positives — title-case words at sentence-initial positions that aren't actually proper nouns, common acronyms the user clearly already pronounces a particular way

For everything the helper missed:

• English place names with silent letters (Worcester, Cholmondeley, Featherstonehaugh)
• Loanwords with retained pronunciation (lingerie, pho, façade)
• Programmer jargon with contested pronunciation (Nginx, kubectl, YAML, GIF)
• Multi-token model names — the helper catches Atari and 2600 separately; merge them by hand into a single candidate Atari 2600

Capture per candidate

For each entry, the table needs:

• Token: surface form as it appears (preserve case)
• Lemma: canonical form for deduping (Siobhán's → Siobhán)
• Category: NAM / CEL / BRA / MOD / ART / DOC / OTH
• Positions: every offset where the lemma occurs in the script
• Original fragment: 5–6 word snippet around the first occurrence, snapped to clause boundaries where possible

Phase 2 — Suggest respellings

Build the pronunciation table in this exact column order:

Respelling rules (full detail in references/respelling-conventions.md):

• Hyphens between syllables: shi-VAWN, never shivawn
• Capitals mark stress: shi-VAWN (stress on second), LLAN-fair (stress on first)
• English orthographic conventions, not IPA: oo for /uː/, ay for /eɪ/, aw for /ɔː/. The model reads letters, not symbols.
• Disambiguate hard consonants: kat not cat if the model is reading the c soft; gohl not goal if the g is going wrong
• Acronyms — letter or word: decide and commit. S-Q-L (letters) or SEE-kwul (word) — both are valid; pick one and tell the user.
• Numerics — write them out: Atari 2600 → Atari twenty-six hundred; ADR-0012 → A-D-R twelve. Lose leading zeros when speaking.

Mark uncertain rows ? in the Respelling column. Don't fabricate.

Phase 3 — Render samples

Find the tool

If the Higgsfield TTS tool has not already been located, search:

tool_search keywords=["tts", "speech", "higgsfield", "eleven"]

If nothing matches, surface the gap to the user via suggest_connectors and pause until the connection is in place.

Render fragments, not words

For each row with status pending:

1. Take the phoneticized fragment (not the isolated word — context matters for prosody; syllable boundaries must land in surrounding speech to be correctly evaluated)
2. Call the TTS tool with the fragment, requesting eleven_v3 (or whatever string the tool exposes for the v3 model)
3. Save the audio to samples/<ID>_<lemma_slug>.mp3 — e.g. samples/P01_siobhan.mp3
4. Fill the Sample column with the relative path

Render in sequence, not in parallel. Higgsfield rate limits, and sequential rendering gives the user deterministic ordering for review.

Present and pause

Present the table with samples linked. Communicate the expected response shape to the user (acceptance, revision, rejection, addition — see Phase 4). Then wait. Do not proceed to Phase 5 without explicit sign-off.

Phase 4 — Iterate

The user's response will mix four kinds of update:

Re-render only revised and newly-added pending rows. Never re-render accepted rows.

Loop until every remaining row is accepted. When the user signs off, proceed to Phase 5.

Phase 5 — Emit phoneticized script

Primary output: script.respelled.txt

The original script with respellings substituted inline. This is the deliverable that goes into Eleven v3.

Substitution rules:

• Every occurrence of every accepted token's lemma → respelling
• Possessives and inflections handled at the surface form, preserving the apostrophe in possessive output as shown in skills/phoneticize/references/respelling-conventions.md lines 133-136: Siobhán's → shi-VAWN's; df12's → dee-eff-TWELVE's
• Sentence-initial capitalization preserved: Siobhán arrived → Shi-VAWN arrived (capitalize the first letter, keep the stress capitals where they were)
• Punctuation preserved exactly
• Multi-occurrence consistency: same lemma → same respelling, every time

Secondary output: script.ssml.xml

For engines that honour <phoneme> tags — Azure TTS, AWS Polly, Google Cloud TTS, older Eleven models. Wrap each substituted token in SSML using the IPA column:

I met <phoneme alphabet="ipa" ph="ʃɪˈvɔːn">Siobhán</phoneme> at the bar.

Skip rows where IPA is empty — emit the bare token, not a malformed tag. The SSML version is best-effort archival; the primary respelling is what the user actually narrates from.

Archive: pronunciation_table.md

The final agreed table, saved alongside the scripts. When the user re-narrates the same content months later, the table lets them reproduce the voicing without re-running Phase 3. Suggest exporting the table as an ElevenLabs PLS alias dictionary too — see references/eleven-v3-notes.md for format.

Output bundle

Final deliverables, presented via present_files:

output/
├── script.respelled.txt        ← primary: feeds into Eleven v3
├── script.ssml.xml             ← secondary: for SSML-aware engines
├── pronunciation_table.md      ← archive of the agreed table
└── samples/
    ├── P01_siobhan.mp3
    ├── P02_df12.mp3
    └── ...

List script.respelled.txt first in the present_files call — that's the deliverable the user will reach for most often.