Kokoro TTS: free narration, if you build the checks

Kokoro is an open text-to-speech model small enough to run on a laptop. Since October 28, 2026 it has voiced nearly all my narration at a cost of zero. It also croaks, stutters on Greek names and reads "live" the wrong way. This page is mostly about catching those before a person does.

What it is

Kokoro-82M has 82 million parameters and an Apache license, so commercial use is allowed. Version 1.0 came out January 27, 2025 with 54 voices across 8 languages. I run it through the kokoro-onnx Python package with two files on disk: the model and a voice pack. It outputs 24 kHz audio. No account, no key, no network call.

If you do not want to host it, the model card notes that hosted Kokoro APIs were selling for under $1 per million characters in April 2025. Cartesia's $49 plan works out to about $39 per million. For high-volume narration that difference is the whole budget.

What I voice with it

  • Narration for a history series on YouTube. Nineteen episodes are scripted or built so far.
  • Voice lines and promo clips for three of my apps.
  • The narration for a hackathon demo film in November 2026.
  • Test audio. When I built a voice-notes tool, I used Kokoro to generate spoken test files so the test suite never needs a microphone.

My narrator is the voice am_liam at speed 1.2. Dialogue characters use am_onyx and am_eric, with am_michael as the backup when one of those misreads a line.

Five failures, and the fix for each

It croaks at the end of short lines

I first generated one line per call. Short lines ended with a pitch drop and vocal fry, and a silence trimmer then cut the tail off the last word. The voice sounded like it was dying at the end of every short sentence.

Fix: synthesize in paragraph groups, so a short punch line is voiced together with its setup. Trim gently and add a fade. Then measure. My checker fails a take if the last word drops more than 6 dB, if there is creak in the final 250 milliseconds, or if the tail is hard-cut.

It garbles hard names inside long passages

"Leonidas" came out as "le-le-onidas" when it sat in a multi-sentence call. The same sentence voiced alone was clean. So grouping, the fix for croak, causes this one.

Fix: any sentence with a difficult proper noun is voiced by itself, then spliced in.

Hyphen respellings make it chant

Writing "THER-MOP-uh-lee" to help with Thermopylae made Kokoro stress every syllable equally. Every name came out chunky.

Fix: give it a phoneme string for the word. Then compare the word's duration to a plain word with the same syllable count. If it runs long, it is chanting.

It picks the wrong reading of heteronyms

"Did he live?" came out rhyming with "five." It has also tripped on "buffet" and on place names like Puget Sound.

Fix: a word list scanned on every script. Each hit gets a forced respelling chosen from context, and goes on a short listen-check list for a human.

A speech-to-text check will pass a stutter

My first quality gate transcribed each take with Whisper and compared it to the script. Whisper is forgiving. It heard a stuttered name and wrote the name correctly, so the stutter shipped. My retry loop made it worse by swapping punctuation until Whisper agreed.

Fix: Whisper checks that no words are missing. It does not get to judge names. Names go to a person.

One more check that has nothing to do with the model

A build once produced an 80.8-second file for seven short lines, about 15 seconds of speech padded with silence and echo. I opened it for review without listening first. Now every audio file gets three automatic checks before anyone sees it: duration against expected, long-silence detection, transcript against script.

What it is not for

Kokoro on my Mac is a batch tool. I have not put it behind a live phone call. A call needs audio streamed in phone encoding within a few hundred milliseconds, and that is the job I keep Cartesia for.

It also has no voice cloning. You get the 54 stock voices.