Dia2
tts · reviewed by Stan K. · Source: data snapshot · data as of 2026-11-12
Apache 2.0 two-speaker dialogue in a single pass — [S1]/[S2] interruptions and reactions without stitching tracks, if you can feed the 1B/2B appetite.
#10 of 16 in Tts
ToolScore breakdown — why 62/100
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
- ▸Two-speaker dialogue ([S1]/[S2]) in a single pass — interruptions and reactions without stitching tracks
- ▸Apache 2.0: the dialogue feature is commercially usable, no NC trap
- ▸Params published (1B / 2B) — the deployment class is knowable in advance
- ▸No quotas, no meters — self-hosted
Where it costs you
- ▸1B / 2B params with VRAM unpublished — a real GPU bill you can't budget precisely (verify pending)
- ▸The dialogue niche is the whole story: nothing else in our snapshot distinguishes it
- ▸No latency number, no arena row — speed and quality unverifiable
Hidden traps — the stuff the pricing page doesn't say
1B / 2B parameters put Dia2 in GPU class by any reading, and the VRAM cell in our hub stays "—" (verify pending). The watts cost model that makes local TTS cheap needs exactly that number — until it publishes, budget a real GPU, not a CPU afterthought.
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: Apache 2.0, 1B/2B params, single-pass two-speaker dialogue (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
≈$0.011/min on the API with Elo 1274 quality — and the local weights are CC-BY-NC, so commercial use
tts-1 verifies at ≈$0.011/min — but gpt-4o-mini-tts bills $12/1M audio output tokens with no publish
The odd one out: edits existing audio by text instruction — fix a word, change the emotion — plus ze
Dia2 review
In-depth analysis · last updated 2026-11-12
Dialogue in one pass
Dia2 does the one thing nobody else in the local table does: two-speaker dialogue — [S1] and [S2] — in a single pass, with interruptions and reactions, without stitching tracks. Every other model here generates one voice at a time and leaves the conversation editing to you. If your product is dialogue — podcasts, animated shorts, interactive voices — that single-pass property is a workflow change, not just a feature.
The economics — GPU class, unknown bill
The local cost model is watts, and Dia2's parameter count (1B / 2B, at least published) says the watts are real. The VRAM cell stays "—" in our hub — verify pending — which means the one number the value-per-VRAM metric needs is missing. That is why a model with a unique, commercially-clean feature sits mid-table: Apache 2.0 earns full marks, the hardware opacity spends them.
Real limits
Structurally none: self-hosted, no quotas, no meters, Apache 2.0 license with no trap in it. The practical limits are deployment-shaped: GPU-class setup with unpublished requirements, no latency figure, no arena row. The feature is the niche; the niche is the product.
How the pillars land
Price/Value 3: the value-per-VRAM verdict — GPU-class parameters with an unpublished bill is the worst cost position among the healthy locals. Quotas 8: no meter exists, docked because a 1B/2B model on unknown hardware makes batch volume an open question. Latency the neutral 5 — no fact, no guess. Training use & voice cloning 8: local inference keeps text on your box; no cloning claims on record. API & self-host 10: Apache 2.0. Feature set 7: single-pass two-speaker dialogue is a genuinely unique capability — scored for the niche it is, not the general-purpose engine it isn't. Track record the neutral 5 — no facts. Ease 4: GPU-class setup with no published requirements is a real deployment bar. Pricing transparency 8: params published, VRAM missing. Support & docs the neutral 5.
Verdict — NICHE, 62/100
Buy it for the dialogue, not as a general engine. If two voices interrupting each other in one render is your product, Dia2 is the only Apache-licensed answer in this table and the feature justifies the GPU. For narration bulk, Kokoro on CPU beats it silly; for cloning, Qwen3-TTS. Dia2 earns its place by doing one thing nobody else does — and pays for it in the hardware column.