Pronunciation Dictionaries & Phoneme Control in TTS

Your narration sounds flawless — until the voice hits your company name and turns three syllables into gibberish. Then it butchers a customer's surname, mangles a drug name, and confidently pronounces "cache" as "cash-ay."

This is the single most common complaint about text-to-speech, and it isn't a bug. Neural voices predict pronunciation from spelling, and spelling lies constantly. English is famously irregular, and proper nouns, brand names, and technical terms fall outside the patterns a model learned.

The fix is control. With a pronunciation dictionary and phoneme overrides, you tell the engine exactly how a word should sound — once — and it stays fixed everywhere. Here's how to do it.

Why neural voices guess wrong

Text-to-speech works in two stages. First, a grapheme-to-phoneme step converts written letters into the sound units the voice will speak. Then the neural model turns those sounds into audio.

Most mistakes happen in that first step. The model has never seen "Voxiven" or "Nguyen" or your API's codename, so it falls back to statistical guesses based on similar-looking English words. Sometimes it's close. Often it isn't.

Homographs make it worse. "Read," "lead," "bass," and "live" each have two valid pronunciations, and the engine picks based on context it doesn't always parse correctly. The International Phonetic Alphabet (IPA) exists precisely because ordinary spelling is an unreliable guide to sound — a problem linguists have documented for over a century (Encyclopaedia Britannica).

The takeaway: you can't fix pronunciation by respelling words and hoping. You need a way to specify sound directly. That's what phonemes give you.

Phonemes and IPA: the vocabulary of sound

A phoneme is the smallest unit of sound that distinguishes one word from another — the difference between "pat" and "bat" is a single phoneme. The IPA assigns each of these sounds a unique symbol, so /ˈkæʃ/ always means "cash" no matter how it's spelled.

Two notation systems dominate TTS work. IPA is the international standard, using symbols like /ʃ/ for "sh" and /θ/ for the "th" in "thin." The other common set is X-SAMPA, an ASCII-friendly encoding of the same sounds for systems that can't handle special characters.

You don't need to memorize the whole alphabet. In practice you'll look up a handful of symbols for the specific names and terms that break. The IPA chart maintained by the International Phonetic Association is the authoritative reference, and most dictionaries now show IPA next to headwords.

Reading a transcription

Take "Voxiven." Break it into syllables — Vox-i-ven — then map each to sound: /ˈvɒks.ɪ.vɛn/. The apostrophe-like mark (ˈ) shows primary stress on the first syllable. Get the stress wrong and even correct sounds feel off, because English listeners key heavily on rhythm.

The <phoneme> tag in SSML

Speech Synthesis Markup Language (SSML) is the W3C standard for controlling synthesized speech, and its <phoneme> element is the tool for overrides (W3C SSML specification). You wrap the word, declare the alphabet, and supply the transcription.

<phoneme alphabet="ipa" ph="ˈvɒksɪvɛn">Voxiven</phoneme>

The alphabet attribute is either ipa or x-sampa. The ph attribute holds your transcription. The text between the tags is what's displayed or logged — the voice ignores it and speaks your phonemes instead.

This solves one-off cases beautifully. Drop it around a tricky surname in an audiobook, a chemical compound in a lecture, or a product name in a promo, and the voice obeys.

If you'd rather not hand-write markup, EchoLive's visual SSML editor lets you build phoneme overrides, breaks, and emphasis through a point-and-click interface, then exports clean SSML underneath. You get the precision of the standard without memorizing the syntax.

From one-off tags to a pronunciation dictionary

Inline tags are perfect for a word that appears once. But if your brand name shows up forty times across a script, tagging each instance is tedious and error-prone — miss one and the illusion breaks.

A pronunciation dictionary — sometimes called a lexicon or substitution list — solves this at the project level. You define a term and its pronunciation once, and the engine applies it everywhere the word appears.

What belongs in your dictionary

Think of it as a living glossary for anything the model can't guess:

Build the dictionary before you narrate a long project, not after. A few minutes of setup saves you from re-recording an entire chapter because the hero's name was wrong throughout. For long-form work like converting a PDF or manuscript to audio, this is the difference between a polished result and an obvious robot.

Substitutions vs. phonemes

There are two ways to fix a word, and they serve different needs. A substitution replaces text with other text — expanding "Dr." to "Doctor" or "St." to "Street." A phoneme override specifies raw sound. Use substitutions for abbreviations and phonemes for anything where spelling can't capture the pronunciation. Good dictionaries support both.

A practical workflow for clean pronunciation

Getting this right is a short, repeatable loop rather than a one-time chore.

1. Scan for landmines. Before generating, read your script for proper nouns, foreign words, jargon, and acronyms. Highlight anything a stranger might mispronounce reading aloud.

2. Look up the sounds. For each term, find its IPA. Reputable dictionaries list it, and pronunciation databases like Wiktionary cover names and loanwords well. When in doubt, transcribe by ear, syllable by syllable.

3. Preview a single word. Don't generate the whole project to test one name. Synthesize just the word with your transcription, listen, and adjust. Moving the stress mark or swapping one vowel often fixes a stubborn case.

4. Promote repeats to the dictionary. Any term appearing more than once goes into your project lexicon so it's consistent across every segment — and reusable on your next project.

5. Do a full listen-through. Even with a dictionary, listen to the final pass. Context-sensitive homographs and stress shifts sometimes slip through, and your ear catches what the eye misses.

This workflow scales from a single social clip to a full audiobook. EchoLive's segment-based Studio editor makes it especially efficient: because voices, styles, and SSML live per segment, you can lock pronunciation in one section and apply the same settings across the rest of a large project with batch operations, without disturbing your pacing.

Why it matters more than it seems

Mispronunciation isn't just cosmetic. It costs credibility. A listener who hears a familiar name mangled instantly clocks the audio as machine-made and lowers their trust in everything else you're saying.

It's also an accessibility issue. For people who rely on audio to consume content, an unpredictable voice is genuinely harder to follow — and clear, correct speech is a core accessibility principle, reflected in guidance like the W3C's Web Content Accessibility Guidelines. Getting names and terms right isn't polish; it's respect for the listener.

The good news is that the effort compounds. Every entry you add to a pronunciation dictionary is permanent infrastructure. Your brand name, your recurring cast of terms, your industry's vocabulary — transcribe them once and every future project starts ahead.

Wrapping up

Neural voices mispronounce words because they predict sound from unreliable spelling. Phoneme overrides using IPA in the SSML <phoneme> tag let you specify exact sound, and a project-level pronunciation dictionary makes those fixes consistent and reusable. Scan, transcribe, preview, promote repeats, and always do a final listen.

If you're producing narration where names and terminology have to land perfectly, EchoLive gives you visual SSML tools, per-segment control, and a full neural voice catalog to make it happen — try it in the playground and hear your trickiest terms come out right the first time.