Turn text into voice with AI

Robotic sound usually comes from the script, not the tool. Long sentences, formal grammar, and unmarked pauses make any engine drone. Fix the words first, then tune the settings.

1. Rewrite in spoken English: use contractions and short sentences.
2. Add SSML pauses (<break time='400ms'/>) at natural breath points.
3. Mark emphasis on the one word that carries each sentence.
4. Drop pace to 90–95% speed for narration; leave 100% for ads.
5. Generate, listen once, and fix only the two ugliest moments.

A 200-page PDF won't fit in one generation, and reading it raw sounds terrible. Split it, clean each chunk, and stitch the audio. The prep matters more than the button-press.

1. Extract clean text — strip page numbers, headers, and footnotes.
2. Split into chapter-sized chunks under your tool's character limit.
3. Run each chunk through an AI to fix hyphenation and OCR errors.
4. Generate audio per chapter with a consistent voice and settings.
5. Concatenate files and add 1-second gaps between chapters.

TTS engines guess at unusual names, acronyms, and numbers, and they guess wrong. You correct them with phonetic spelling or SSML tags, not by hoping.

1. List every proper noun, acronym, and figure in your script.
2. Generate a quick test of just those words to hear the failures.
3. Respell mispronounced names phonetically (e.g. 'Nguyen' → 'Win').
4. Wrap acronyms in <say-as interpret-as='characters'> for letter-by-letter.
5. Force number formats with <say-as interpret-as='cardinal' or 'date'>.

Kokoro 82M is small enough to run locally, which means free unlimited generation once it's set up. It needs Python and a bit of patience, not a monster GPU.

1. Install Python 3.10+ and pip on your machine.
2. Install the Kokoro package and download the v1.0 model weights.
3. Load a voice pack and confirm the sample rate you want.
4. Feed it a short test sentence and check output quality.
5. Wrap it in a loop to batch-process your full script.

Don't run English text through a Spanish voice — translate first, then pick a native voice for each language. Machine translation plus a native TTS voice sounds far better than one voice faking an accent.

1. Translate your script per target language with a strong AI model.
2. Have a native speaker or AI check the translation for tone.
3. Pick a native voice profile in Azure or Qwen for each language.
4. Generate separately — never reuse one voice across languages.
5. Spot-check pronunciation of brand names in every version.

Old TTS glued together recorded phonemes — that's the flat, choppy computer voice you remember. Neural TTS predicts the whole waveform, so it flows with real intonation and rhythm.

• Neural models learn prosody — the rise and fall of real speech.
• They handle pauses, emphasis, and emotion instead of monotone.
• They mispronounce fewer rare words but still need correction.
• They cost more compute, so cloud tools charge per character.
• The gap between neural and human is now small for short clips.

SSML is markup — like HTML for speech. You wrap parts of your text in tags to control how the AI says them, instead of accepting its default guess.

• <break> inserts a pause exactly where you want breathing room.
• <emphasis> stresses a word so it doesn't get swallowed.
• <prosody> changes speed, pitch, and volume mid-sentence.
• <say-as> forces numbers, dates, and acronyms to read correctly.
• <phoneme> spells out tricky pronunciations letter-sound by sound.
• Not every tool supports every tag — Azure has the fullest set.

Usually yes, but it depends on the tool's license, not on general AI law. Azure and most paid TTS grant commercial rights on generated audio. Kokoro's Apache license lets you sell too.

Where you get burned is voice cloning of a real person without consent, or free-tier tools that ban commercial use. Read the license before you monetize, not after.

Use Kokoro locally — it's free and unlimited, so cost is zero and you can regenerate as many times as you need. Write your script tonight, clean it with an AI for spoken flow, then batch-generate.

Pick one voice and stick with it for consistency. Don't chase perfection at midnight — get a clean take of the whole thing, fix the two worst lines, and export. You'll have audio before your alarm goes off.

Don't try to make 40 unique voices from scratch. Group your NPCs by type — gruff guards, cheerful merchants, ancient wizards — and assign a handful of base voices. Then vary pitch and pace per character using prosody settings to multiply them.

Qwen and Azure both offer many voice profiles; Kokoro has fewer but runs free for bulk lines. Build a spreadsheet mapping each NPC to voice + settings so every line stays consistent across your whole game.

You want automation, not hand-generating each post. Wire a TTS API to your publishing flow so new posts get audio automatically. Azure Neural TTS has a clean API and per-character pricing that's cheap for text-length articles.

Clean the post text first — strip image captions and code blocks that read as gibberish aloud. Pick one warm, mid-pace voice and keep it site-wide so listeners get a consistent experience.

Azure is the safe workhorse: huge voice library, rock-solid SSML, reliable at scale. Qwen shines on expressive, emotional delivery and handles Chinese and English especially well. Pick by what your content needs most.

Cloud gives you the best voices with zero setup but charges per character and needs internet. Local Kokoro is free forever and private, but the voice range is smaller and you handle the setup yourself.

For flat, functional narration — course modules, IVR menus, blog readers — yes, AI already does the job cheaper. That work was never glamorous and it's mostly gone to AI.

But AI still can't nail true performance: comedic timing, a character breaking down in grief, the choices a real actor makes on the fly. High-craft voice work is safe for now. The middle is where it's hollowing out.

Trust the tool for clarity, but check every generation before it ships — AI voices mispronounce brand names, botch numbers, and occasionally add a weird inflection that changes your meaning. Listen to the full clip once, don't just skim the waveform.

The bigger risk is sameness: if you use a stock voice, so do a thousand other brands. Pick a distinctive voice, keep it consistent everywhere, and you'll build recognition instead of blending in.