Skip to content
Back to all essays
ProjectsAnnouncements

Announcing Ghost Narrator: Self-Hosted AI Narration Stack

By George PuEssay7 min read

Published Updated

Announcing Ghost Narrator: Self-Hosted AI Narration Stack
Listen to this essay
0:000:00
AI-generated voice

Listen to article 0:00 / 0:00
Speed

Historical launch context — October 11, 2026: This is the March 31 Ghost Narrator announcement. Publishing-volume, narration coverage, hardware/setup estimates and planned upgrades describe the launch account, not current service guarantees. The code’s MIT license does not replace model-license and voice-consent checks. The repository is the project reference; current maintenance has not been verified for this note. Local generation can still involve publishing/upload integrations.

At launch I reported about 200 blog posts a month and intended narration for each post — you can listen instead of read.

The obvious solution was ElevenLabs. Best-in-class voice cloning, simple API, great output. $330/month for the scale we needed.

We almost signed up. Then we tried something else.

We ran an open-source model on a laptop.

I described a local LLM as Qwen 3.5 14B for script rewriting and Qwen TTS for voice generation. Exact checkpoint identifiers were not supplied in this launch account. The generation steps ran locally; website publishing used integrations, and compute capacity still limited throughput.

It sounded fine. Not perfect. Fine. Good enough that nobody listening would think twice.

So we built it into a system, and today we're open-sourcing it.

It's called Ghost Narrator.

What It Does

Ghost Narrator turns written blog posts into narrated audio — automatically.

No manual steps. No copy-pasting into a TTS tool.

Here's the flow:

  1. A new article gets published
  2. An automated workflow triggers
  3. A local LLM (Qwen 3.5 14B) rewrites the article into a natural narration script — removing headers, links, formatting artifacts, and anything that sounds wrong when spoken aloud
  4. An open-source voice cloner (Qwen TTS) generates audio using a cloned voice
  5. The audio file gets attached to the post

The generation steps can run locally. Attaching the finished audio to a hosted blog still involves its publishing integration; this launch description does not establish that every byte stays on the machine.

How It Works

Ghost Narrator has two core pieces working together.

The first is a local LLM — we use Qwen 3.5 14B running through Ollama. When a new article comes in, the LLM rewrites it into something that sounds natural when spoken.

Blog posts are written for eyes — they have headers, bullet points, hyperlinks, parenthetical asides. None of that works as audio.

The LLM strips all of it out and produces a clean narration script that flows like someone telling you a story.

The second is the Qwen TTS model we described for voice generation. We used sample audio of my voice. Generation ran locally in that setup; the resulting resemblance was a listening judgment, and publishing the audio was a separate integration.

The cloned voice reads the narration script, and out comes an audio file.

Ghost Narrator ties these together into a single pipeline. Article goes in, audio file comes out. That's it.

The Cost Comparison

This is the part people keep asking about.

ElevenLabs was quoted at $330/month for the Scale plan at publication, or $3,960 for twelve months at that monthly charge. This account did not include a dated invoice, the full allowance and billing terms, or an accepted-output comparison. That quoted hosted-service cost is historical, not current pricing or proof of equivalent workloads.

Ghost Narrator uses compute you supply. Electricity is one cost; hardware, setup, licensing, review, maintenance and publishing infrastructure also matter. The locally run model avoids a hosted TTS provider’s per-word charges, but compute still has capacity limits. Audio publishing uses configured external services. Setup time and ongoing review depend on the installation; this account does not establish an hour-or-two setup guarantee.

At launch I preferred ElevenLabs for some aspects of naturalness and emotion. That was a listening impression, not a blind industry-wide quality benchmark.

If you're producing audiobooks or premium podcasts, you probably still want a commercial solution.

At launch I hoped later models would improve the output. I had not established a six-month quality forecast; each version would need listening and compatibility checks.

A modular pipeline can make model changes easier, but a new model may need different hardware, integration work, license checks and review. Better output is an outcome to test, not an automatic consequence of swapping.

The other lever is hardware. Bigger machines can run bigger models. A 16GB M1 MacBook runs the lightweight voice models fine.

A Mac Mini M4 Pro with 64GB can run the larger, higher-quality models that produce noticeably better output. Same pipeline, same setup — just more room for the model to work with.

We plan to keep improving the voice output as better models get released and as we test on better hardware, and we'll document every upgrade on this blog.

We planned to test future model upgrades. Better output or unchanged integration is not guaranteed; each swap needs compatibility, license and listening checks. Those were upgrade goals, not verified outcomes or a promise that integration remains unchanged.

For our reported launch workload we preferred self-hosting. We did not publish a complete like-for-like total-cost or blind quality comparison that proves it wins for another team.

We avoided a hosted TTS provider’s per-word charges for local generation. Batch throughput still depended on our hardware, and publishing the completed audio to the website involved external infrastructure.

What Hardware You Need

We run this on a MacBook Pro M3 with 36GB RAM. That's more than enough.

The Mac Mini M4 Pro (64GB) is the sweet spot if you want a dedicated narration server. It runs everything with room to spare. We're moving our production setup to one.

Minimum specs: an Apple Silicon Mac (M1 or later) with 16GB+ RAM, or a Linux box with a decent GPU (RTX 3090 or better). About 50GB of free disk space for models.

This isn't a cloud GPU situation. This runs on hardware you'd buy once and keep on your desk.

Who This Is For

Content teams publishing at scale who are tired of per-word TTS pricing. Developers who want a working example of a self-hosted LLM + voice cloning pipeline running locally. Anyone who looked at their ElevenLabs bill and thought "there has to be another way."

If you publish 10 blog posts a month, ElevenLabs is probably fine. If you publish 100+, the math changes fast.

What This Isn't

This isn't a polished product. It's an open-source project released as-is.

There's no one-click installer. There's no support team. There's no guaranteed uptime. You'll need to be comfortable with Docker, command-line tools, and reading a README.

We built this for ourselves. We're releasing it because people kept asking how we did it. If it's useful to you, great. If it breaks, you get to keep both pieces.

The project code was released under MIT, subject to its conditions. Model weights, voice rights and other dependencies require their own checks.

Try It

Listen: Article archive links to the archive; look for a player on the particular article you want to hear. This launch post does not establish universal narration coverage.

When we publish a new article, it'll be entered in a queue for narration. The latest blog posts may not have narration right away.

Clone it: https://github.com/getsimpledirect/ghost-narrator

Run it: The README has everything you need. Prerequisites, setup, configuration.

What's Next

Ghost Narrator is the first project we're open-sourcing from SimpleDirect. More are coming.

At publication we wanted more models, tools and generation infrastructure to run on hardware we controlled. That was a direction, not evidence that all API dependencies, subscriptions or outside infrastructure had been eliminated. We intended to publish further replacements and accounts as we tested them.

If that interests you, follow along:

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.