Dia2
Nari Labs' second-generation dialogue TTS: streaming two-speaker conversation synthesis in one take — 1B and 2B Apache-2.0 checkpoints.
About
Dia2 is the follow-up to Nari Labs' viral Dia 1.6B dialogue TTS, and it doubles down on the same idea: generate ultra-realistic multi-speaker dialogue directly from a transcript in a single pass. Where Dia1 rendered a full conversation offline, Dia2 is a streaming dialogue model — it synthesizes from partial text in real time, with [S1]/[S2] speaker tags trading lines with natural conversational turn gaps. Each speaker can be conditioned on a voice prefix (a reference clip plus its transcript) for tone and identity control. Under the hood it's a delayed-streams model: a 28-layer decoder reads the script word by word while a 4-layer depformer emits Kyutai Mimi's 32 codebooks per frame, decoded to 24 kHz audio. Nari Labs ships 1B and 2B checkpoints under Apache 2.0, with the 2B model on Hugging Face (nari-labs/Dia2-2B) and an active community porting it to Apple silicon via MLX. It's English-only for now, and Nari Labs' quality claims rest on their own side-by-side demos against ElevenLabs Studio and Sesame CSM-1B — compelling, but not independent benchmarks.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- No published benchmarks; Nari Labs publishes side-by-side comparison demos vs ElevenLabs Studio and Sesame CSM-1B (not independent)
- No independent benchmark replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: Nari Labs open release (Nov 2025); Apache-2.0; 1.2k GitHub stars on the dia2 repo; HF checkpoints with 5k+ downloads; community MLX port. Strengths: streaming two-speaker dialogue in one take, voice-prefix conditioning per speaker, nonverbal cues (laughter, coughing), fully open license. Docked for: (1) English-only and no published benchmarks — quality evidence is Nari Labs' own demo comparisons; (2) young project from a tiny team, so maintenance and scaling are unproven. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Nari Labs Dia2: 1B + 2B checkpoints, streaming two-speaker dialogue TTS in one take
- · Repo created 2025-11-17, Apache-2.0, 1,182 stars; HF Dia2-2B with 5k+ downloads; community MLX port for Apple silicon
- · Delayed-streams model: 28-layer decoder + 4-layer depformer emitting Kyutai Mimi's 32 codebooks, 24 kHz out
- · [S1]/[S2] speaker tags; voice-prefix conditioning (clip + transcript); nonverbal cues; English only
- · No published benchmarks — Nari Labs' own demo comparisons vs ElevenLabs Studio and Sesame CSM-1B