Skip to content
Back to all essays
Own Your TechProjects

We Swapped Our TTS Model. The Older, Free One Won.

By George PuEssay5 min read

Published Updated

We Swapped Our TTS Model. The Older, Free One Won.
Listen to this essay
0:000:00
AI-generated voice

Listen to article 0:00 / 0:00
Speed

Model and comparison context — October 11, 2026: This is our April 8 model-swap account and listening preference. It does not include exact tested checkpoints/revisions, matched audio samples, blind listener ratings, latency or cost logs. For reference, the official Qwen3-TTS-12Hz-1.7B-Base card lists Apache 2.0, while Voxtral-4B-TTS-2603 lists CC BY-NC 4.0. Those cards identify specific artifacts; they do not prove which revisions we tested. The project reference is Ghost Narrator; current maintenance and narration coverage are not verified by this old post.

When we launched Ghost Narrator, we used Fish Speech for voice cloning. It worked. The voice quality was past the threshold of "nobody notices." We were happy with it.

Then we had a licensing problem.

Fish Speech's license is restrictive for commercial use.

We were publishing 200 narrated blog posts a month on a model we couldn't fully commercialize. That needed to change.

So we tested alternatives. The obvious move was Mistral's new TTS model — newer, from a well-funded lab, getting attention on Twitter. We tried it.

The license was restrictive too.

Then we tested Qwen TTS.

Apache 2.0. Permissive for commercial use subject to the exact artifact’s license conditions, including required notices; voice rights and other dependencies still need checking.

And here's the part that shocked us: we preferred the voice quality in our listening tests.

Not marginally. Not "if you listen carefully." Significantly.

Qwen published this model three months before Mistral published theirs.

Older model. More permissive license. Better output.

Every assumption we had was wrong.

  • We assumed newer meant better.
  • We assumed more restrictive meant more advanced.
  • We assumed the well-funded Western lab would outperform.

None of that held up.

What Changed

The swap was straightforward.

Ghost Narrator's architecture is modular — the TTS model is one component in the pipeline.

We pulled Fish Speech out, dropped Qwen TTS in.

Same pipeline, same workflow, same hardware.

The narration scripts and voice generation ran locally in our setup. Publishing the resulting audio to the website involved configured external services; local generation is not an end-to-end no-data-leaves guarantee.

The only difference is the model producing the audio.

If you cloned Ghost Narrator before the swap, updating is a model change. The architecture doesn't care which TTS model you run.

Listen to It

At publication we were moving narration to Qwen TTS. A working player on a particular article is evidence of that audio, not proof that every article has narration or uses the same model revision.

What This Means

A year earlier, I was treating paid APIs such as ElevenLabs as the default for the voice quality I wanted. That was my starting assumption, not a comparison establishing the best TTS option across the market.

In our April listening comparison, we preferred the locally run Qwen option to the alternatives we tried. Its permissive model license did not make hardware, review and publishing free.

For the narration we listened to, the locally run option suited us better; this comparison does not rank the whole paid and open-weight TTS market.

I saw a risk for businesses differentiated mainly by model access. The 18-month/most-companies prediction was speculation, not a measured failure forecast.

Our experience made me question the assumption that a newer or more restrictive model would necessarily suit us better.

I did not measure a two-year or one-year quality lag across model releases.

The older model was the one we preferred for this narration workflow; that does not establish a negative market-wide lag or an objective benchmark win.

For our workflow, access to a particular model looked less durable as a differentiator once we found an acceptable replacement.

Our swap made that particular dependence less valuable to us; other products can still provide quality, reliability, support and workflow value.

I liked ElevenLabs’ voice quality at the time. That preference and our Qwen swap are different observations; neither provides an industry-wide ranking.

The Qwen artifact we described had more permissive terms for our intended use, and we preferred the output we heard. I cannot establish the developers’ competitive intent or general superiority from that experience.

This is why we built Ghost Narrator the way we did.

Modular. Local. Model-agnostic.

We didn't bet on Fish Speech being the best forever. We bet on the architecture being right and the models being swappable.

That bet paid off in three months instead of three years.

For our project, the ability to change a model without rebuilding the whole workflow proved useful.

I want to preserve that flexibility. Paid services may still be worth their cost for quality, reliability, support or workflow features; our swap does not predict which companies will survive or when another service will become free.

We're going to keep documenting every swap. Every time a paid tool gets replaced by something we run locally, we'll write about it here.

Not because we're anti-commercial software. Because the math is changing faster than people realize, and I'd rather show it than argue about it.


Ghost Narrator is open-source: github.com/getsimpledirect/ghost-narrator

Follow along: @TheGeorgePu

Share this:

New essays and letters, by email

Subscribe if you would like to hear from me when I have something worth sending. Everything here remains free to read.

Optional. Unsubscribe anytime.