You Don't Need a Paid TTS API
5 local models, $0 per minute the number that matters here
By Stan K.
Piper, MeloTTS, Kokoro-82M, Dia2 and Qwen3-TTS are all permissively licensed and cover UI voice, prototyping, personal audiobooks and offline embedding at $0 per minute. The three cases where paying still wins — top-brand quality (Elo 1338), licensed cloning, hosted latency.
Most TTS invoices pay for quality headroom the workload never uses. If the voice is UI feedback, a prototype, a personal audiobook or an offline feature inside your app, five permissively licensed local models already cover the job — at $0 per minute, no quota, no rate limit, no terms-of-service risk. This is the honest routing table, including the three cases where paying genuinely wins.
The five models and the jobs they own
| Model | License | The job it owns |
|---|---|---|
| Piper | MIT | CPU-only by design — the lowest-end hardware that can still speak |
| MeloTTS | MIT | CPU-capable multilingual synthesis, easy to embed |
| Kokoro-82M (82M) | Apache 2.0 | Production narration quality on CPU — our measured workhorse |
| Dia2 (1B / 2B) | Apache 2.0 | Two-speaker dialogue ([S1]/[S2]) in a single pass — interruptions and reactions without stitching tracks |
| Qwen3-TTS (0.6B / 1.7B) | Apache 2.0 | 3-second rapid voice clone without a paid API |
Every license above was read from the LICENSE file — which matters, because the “free TTS model” lists are full of traps: Fish Speech 1.5’s weights are CC-BY-NC (non-commercial) and Breeze TTS 2’s weights forbid commercial use even self-hosted. The full rogues’ gallery is in Seven TTS Fine-Print Traps.
What “good enough” actually covers
- UI voice — notifications, prompts, micro-feedback. Nobody needs arena-winning prosody to hear “build finished.” Piper on a potato does this job forever, offline.
- Prototyping and iteration. When you’re rewriting the script ten times an hour, per-character billing is a tax on editing. Local regeneration is free — and regeneration is the hidden multiplier in every TTS bill, as the per-minute math shows.
- Personal audiobooks and long-form for yourself. Volume is exactly where paid APIs hurt and local shines — the economics that made us run our own audiobook and narration pipelines for years (what we learned).
- Offline embedding in an app. A voice that must work without network, without an API key, without a per-call bill — that’s not a cheaper cloud, it’s a different product. Local is the only offer.
The three cases where paying wins — honestly
This article is not “local always wins.” It doesn’t.
- Top quality for a public brand. The AA Speech Arena puts Eleven v4 at Elo 1338 ($80/1M chars, v4 Turbo at 1337 for $40/1M) — and if this voice is the headline of a podcast with thousands of listeners, that gap is your product. Notably, the arena’s value row is also a reason to pay carefully: Simba 3.5 sits in the top 5 at Elo 1293 for $6.58/1M chars.
- Cloning a specific voice with the rights handled. Qwen3-TTS’s 3-second clone covers experimentation; cloning a real person’s voice for commercial work is as much a legal product as a technical one, and vendors selling licensed clones carry that weight so you don’t.
- Hosted latency with an SLA. Eleven v4 Turbo’s ~100 ms median is an official number someone contractually stands behind. If your product is a live conversation, “under 200 ms, in the cloud, or we answer for it” is a purchase, not a nicety.
If the voice is content — heard by an audience, attached to a brand, cloned from a real person — budget for paid. If the voice is infrastructure — UI, prototypes, personal listening, offline features — the five models above end that bill permanently. When in doubt, prototype on Kokoro for a week: the cost of finding out is zero.
The complete price-per-finished-minute comparison, every vendor’s traps and the full local leaderboard live in the TTS hub; the hardware to run all of this is ranked in Local GPU AI and One GPU Beats Ten Subscriptions.