MeloTTS
tts · reviewed by Stan K. · Source: data snapshot · data as of 2026-11-12
MIT-licensed, CPU-capable multilingual synthesis that is easy to embed — the mid-table local pick when you need languages, not heroics.
#8 of 16 in Tts
ToolScore breakdown — why 70/100
10 criteria × 10 points. Low pillars are visible on purpose — that's the point of a score you can audit. Methodology →
What it does well
- ▸CPU-capable multilingual synthesis — the language story most local rivals in our table don't have
- ▸Easy to embed — the integration path is part of the pitch, not an afterthought
- ▸MIT license: no commercial restrictions, no NC trap
- ▸No quotas, no meters — self-hosted volume without fine print
Where it costs you
- ▸Neither params nor VRAM are published in our snapshot — the watts bill is verify pending
- ▸"CPU-capable" is weaker than Piper's "CPU by design" — a GPU may still be the sane deployment
- ▸No latency number and no AA Speech Arena row — speed and quality both sit unverifiable
Hidden traps — the stuff the pricing page doesn't say
Params: —. VRAM: —. Our hub publishes hardware requirements only when a source confirms them. MeloTTS is CPU-capable per its model card line, but sizing a deployment — the whole point of the local watts cost model — is guesswork until those cells fill.
Know the limits before you pay: read the traps above, then decide with eyes open.
Affiliate link where available — never affects the score.
- internal ToolScore TTS hub — local table: MIT license, CPU-capable multilingual synthesis, easy to embed (licenses from LICENSE files and model cards) verified
Data as of 2026-11-12. Prices and limits change monthly — report anything stale. How we verify: methodology.
Closest alternatives
The top of the AA arena — Elo 1338 (v4) and 1337 (v4 Turbo) with ~100 ms median latency — and the pr
≈$0.011/min on the API with Elo 1274 quality — and the local weights are CC-BY-NC, so commercial use
MIT-licensed and CPU by design — the lightweight workhorse for the machine where there is no GPU at
MeloTTS review
In-depth analysis · last updated 2026-11-12
The multilingual middle
MeloTTS occupies the quiet middle of the local table: CPU-capable multilingual synthesis, easy to embed, MIT-licensed. It is the pick when the requirement is languages rather than heroics — Piper has no multilingual story on record, Kokoro's story is English-forward production narration, and the cloning/dialogue specialists (Qwen3-TTS, Dia2) solve different problems.
The economics
Watts, not minutes — but two cells are blank: our hub table carries "—" for both params and VRAM, because we publish hardware requirements only when a source confirms them. "CPU-capable" says a GPU isn't mandatory; it does not say a GPU isn't better. Until the blanks fill, the watts bill — the entire local cost model — is verify pending. That opacity is the main reason this rows sits below Piper despite the stronger feature claim.
Real limits
Structurally none: self-hosted, no meter, no quotas, MIT license with zero commercial restrictions. The practical limits are the unverifiable ones: no latency number, no arena row, no parameter count. What is on the record — multilingual synthesis and easy embedding — is exactly what its rivals don't claim.
How the pillars land
Price/Value 7: electricity only, but the hardware bill is uncomputable with params and VRAM both unpublished — the lowest value score among the healthy MIT/Apache locals for that reason. Quotas 9: self-hosted, no meter exists. Latency the neutral 5 — no fact, no guess. Training use & voice cloning 9: local inference, nothing leaves the box; no cloning on the record. API & self-host 10: MIT, embed anywhere. Feature set 6: multilingual synthesis is a real feature claim the rest of the mid-table lacks — scored modestly because it's the only one. Track record the neutral 5 — no facts. Ease 6.5: "easy to embed" is a published claim, docked because sizing the hardware is guesswork. Pricing transparency 8.5: license published, both hardware cells blank. Support & docs the neutral 5.
Verdict — GOOD, 70/100
A sensible default for multilingual narration on a CPU-class box you already run. If your hardware has no GPU at all, Piper wins; if you need cloning, Qwen3-TTS wins; if you need neither, MeloTTS does the job with the cleanest license there is. Ask the vendor to publish the specs sheet — two blank cells are all that keep it out of the 80s.