What Years of Local TTS Taught Us
3 production lines, 0 TTS bills the number that matters here
By Stan K.
Three production lines — English narration on Kokoro-82M, instruction-based audio editing on Step-Audio-EditX, and a Bulgarian audiobook pipeline mastered to -19 LUFS — run on local models with zero TTS invoices. The honest limits: cloning, polyglot quality, latency.
We stopped pricing TTS APIs for most of our work years ago — not out of ideology, but because three production lines kept running without a single voice invoice, and the invoices we did model never survived contact with the volume. This is what those years actually taught us, including the parts local TTS still can’t do.
The workhorse: Kokoro-82M
Most of what people call “AI voiceover” is English narration, and Kokoro-82M (Apache 2.0, 82M parameters) covers it embarrassingly well for its size. It is our measured workhorse: on our RTX 3060 box it runs CPU-only, at roughly 20× faster than real time — the GPU stays free for the jobs that actually need it. An 82M model is a rounding error next to today’s LLMs, and it still voices production narration that ships. The lesson: for a huge share of TTS work, the quality bar is “clear, consistent, pleasant for an hour of listening” — not “wins a blind arena vote.”
The editor: Step-Audio-EditX
The second line is the strange one. Step-Audio-EditX (Apache 2.0, 3B, ~6–8 GB quantized, wants a 12 GB+ GPU) doesn’t just synthesize — it edits existing audio by text instruction: fix one mispronounced word, change the emotion of a line, keep everything else bit-identical. That is a different product from regeneration. On a paid API, “fix one word” means re-synthesizing and re-billing the whole segment and hoping the new take matches; here it means an instruction and a diff. It also does zero-shot TTS, which is why it lives inside our video pipeline — the same box renders the audio and the visuals.
The long haul: an audiobook is a pipeline, not a model
Our longest-running local TTS project is a multi-volume audiobook line in Bulgarian, and the model turned out to be the easy part. The real engineering is everything around it: text extraction from PDFs, machine translation, chunking at natural pause points, synthesis with a local voice, then an ffmpeg mastering chain that ends at a -19 LUFS delivery target. Raw WAV masters are archived so the mastering chain can be re-applied when a target changes — re-mastering must never mean re-synthesizing. Volume after volume, the cost per finished hour is the same: electricity.
Long-form voice is dominated by pipeline concerns — chunking, mastering loudness, resumability, keeping raw masters — not by model choice. Teams that bill by the character feel this twice: once for the synthesis, once for every regenerated take. We covered that multiplier in The Real Cost of a Finished Minute.
Where local honestly stops
Three limits, stated plainly:
- Voice cloning. Neither Kokoro nor Step-Audio-EditX clones a voice from a sample. The Apache-licensed option with a confirmed 3-second rapid voice clone is Qwen3-TTS (0.6B/1.7B) — when a job needs a specific human’s voice with the rights to match, that’s the local path, and if it isn’t good enough, that’s the moment the paid-API conversation starts.
- Polyglot quality. Quality is not uniform across languages. Kokoro is our English workhorse; the Bulgarian audiobooks run a different local voice. We have yet to find one open model that is simultaneously the best at English narration and Bulgarian fiction — plan on per-language voices, not one model to rule them all.
- Latency, by construction. Local inference has no network hop — no round-trip, no cold start, no rate limit. The official cloud benchmark to beat is Eleven v4 Turbo at ~100 ms median latency (their docs), and that’s a genuinely impressive number. But for interactive, on-device or offline work, “the model is already on your machine” is a latency tier no wire can reach.
What we’d tell anyone starting
Buy nothing first. Put Kokoro on whatever CPU you own and find out what percentage of your voice work it already covers — for prototyping and narration, that percentage tends to be uncomfortable for vendors. Then read You Don’t Need a Paid TTS API for the full routing table, and One GPU Beats Ten Subscriptions for the stack this slots into. The complete per-minute price comparison — including what the paid options cost when you do need them — is in the TTS hub.