Feature request: IPA / phoneme input support + improved pronunciation consistency

Problem

Chatterbox currently only accepts plain grapheme (spelled) text input, with no way to specify pronunciation directly, and pronunciation/delivery can vary noticeably across separate generations of the same input. For projects with proper nouns, foreign-language loanwords, or invented terminology, this creates real production problems:

1. No phoneme control. Certain words are mispronounced, with no way to correct them other than respelling the input text phonetically (e.g. "Hveem" → "Ha-veem", or inflected Norwegian forms like "berserkerne" → "bersærkerne") and locking in whichever variant sounds acceptable through trial and error. This doesn't scale well across large projects (e.g. audiobook production) where the same problem word occurs dozens of times, and a respelling that works for a word's base form doesn't reliably carry over to its plural, compound, or possessive forms — each inflected variant has to be separately tested and locked down by ear.

2. Inconsistent pronunciation across generations of identical text. Even with the same input text and same settings, the same word can render differently between generation sessions — sometimes correctly, sometimes not — making it hard to trust a "fix" will hold if that line ever needs to be regenerated later (e.g. after an edit elsewhere in a long-form project). This also shows up as broader delivery/style inconsistency: the same voice can sound noticeably more or less dramatic/expressive between sessions even at fixed exaggeration/cfg_weight/temperature values, which is disruptive when patching or extending a long-form work (audiobooks, serialized content) where consistency with previously generated audio matters as much as per-line quality.

Use case (concrete example)

Producing a bilingual audiobook (EN/Norwegian) with proper nouns and Old Norse-derived vocabulary. Distributor QC flagged specific words as "pronounced incorrectly or inconsistently." Fixing this required:

1. Manually testing multiple respelling variants per word per language
2. Locking in a specific grapheme substitution as the "TTS input" while keeping the original spelling in the actual manuscript
3. Repeating this process separately for every inflected form of the same root word
4. Discovering that patches generated in a later session sounded stylistically different from the original full-book render even when using nominally matching settings, requiring a separate investigation into which generation parameters/model variant were actually used originally

Both a phoneme input path and better generation reproducibility would meaningfully reduce this kind of rework.

Suggested features

a) Phoneme/IPA override — either:

• Inline markup similar to SSML <phoneme> tags, or
• A supplementary pronunciation-dictionary/lexicon file (word or word+context → IPA) applied at inference time, similar to custom lexicon support in other TTS projects (e.g. Coqui TTS, festival-based systems)

b) Deterministic/reproducible generation mode — a documented way to reproduce the exact same output (pronunciation and delivery style) for the same input + seed + parameters, and clearer documentation of which parameters affect delivery consistency across the Standard/Multilingual/Turbo variants (e.g. Turbo currently ignores exaggeration/cfg_weight — this kind of variant-specific behavior difference would be useful to have called out explicitly, since it's easy to unknowingly mix variants across a long project and get inconsistent results).

Even limited support (a lexicon override for exact-match tokens, plus better seed/parameter documentation) would meaningfully help projects with heavy proper-noun content or long-form/multi-session production, without requiring full IPA parsing or a guaranteed-deterministic model.

Environment

• Chatterbox (open-source, character-based TTS)
• Use case: multilingual audiobook production, Standard/Multilingual/Turbo variants