# Dia / Dia2

Nari Labs' dialogue-focused TTS — generates ultra-realistic
multi-speaker conversation in a single pass — served via
[devnen/Dia-TTS-Server](https://github.com/devnen/Dia-TTS-Server), the
same actively-maintained OpenAI-compat wrapper author as our
[chatterbox](../chatterbox/) stack. The wrapper carries multi-model
support for the whole Dia family:

| Model | HF repo | Notes |
|---|---|---|
| Dia 1.6B | [nari-labs/Dia-1.6B](https://huggingface.co/nari-labs/Dia-1.6B) | original, dialogue in one pass (in-image default) |
| Dia2-1B | Nari Labs Dia2 family | streaming, lower latency |
| Dia2-2B | Nari Labs Dia2 family | highest quality |

Dia2 was released 2025-11-19 ([nari-labs/dia2](https://github.com/nari-labs/dia2)).

**Server:** irv-ml1 (Irvine, WireGuard-only)
**Port:** 8200 (container listens on 8003)
**GPUs:** pins to device 0 (RTX 3090) by default; ~7 GB VRAM at BF16
**Image:** `local/dia:v1` — built locally from a pinned git SHA of the
wrapper repo via docker buildx's git URL context
**Upstream wrapper:** [devnen/Dia-TTS-Server](https://github.com/devnen/Dia-TTS-Server) (MIT)
**Upstream model:** [nari-labs/dia](https://github.com/nari-labs/dia) / [nari-labs/dia2](https://github.com/nari-labs/dia2) (Apache-2.0 weights)

## Why this stack exists

Fills the **dialogue-scene** slot none of the other TTS own cleanly.
Dia generates multi-speaker turn-taking in one pass with inline `[S1]`/
`[S2]` speaker tags and nonverbal cues — `(laughs)`, `(coughs)`,
`(sighs)`, `(clears throat)` — directly in the prompt. That's a
different shape from the single-speaker engines:

- **Fish S2-Pro / IndexTTS-2 / Chatterbox** are excellent single-voice
  readers (rich emotion, cloning) but you'd have to stitch turns
  yourself.
- **VibeVoice** does long-form multi-speaker but is podcast/narration
  shaped, not fast turn-taking with nonverbals.
- **Dia** is the one built for *scene dialogue* — the
  multi-character-storytelling case skaldsong is aimed at.

OpenAI-compatible (`POST /v1/audio/speech`), so skaldsong can target it
by base-URL once we add a `dia` engine option to its router.

## Deploy

```bash
# from this workstation (irv-ml1 is WG-only — routes via ana-wg):
scripts/deploy-stack.sh irv-ml1 dia
# then on irv-ml1, first run builds the image from the pinned SHA:
#   docker compose up -d --build
```

First boot pulls the checkpoint (~6-10 GB) into `DIA_CACHE_DIR` and can
take several minutes; the healthcheck's 600 s `start_period` covers it.

## Deployment shape (as of 2026-05-31)

This stack now runs **two fixed-model instances** from `local/dia:v2`
(the dia2-capable image — see [`dia2-image/Dockerfile`](dia2-image/Dockerfile)):

| service | model | port | notes |
|---|---|---|---|
| `dia2-2b` | nari-labs/Dia2-2B | 8200 | highest quality |
| `dia2-1b` | nari-labs/Dia2-1B | 8202 | streaming / lower latency |

The wrapper is single-model and ignores per-request model selection, so
one fixed instance per model is the only way to offer both as real
asset-engine choices. Legacy Dia 1.6B was retired. `local/dia:v2` is built
in two stages: upstream wrapper → `local/dia:v1`, then `dia2-image/` layers
in the dia2 package + its deps. Each instance pins its model via a mounted
`/opt/docker/conf/dia2-*/config.yaml`.

## Voices — stabilizing the random-voice behavior

Dia2 samples a **random speaker (random gender) per generation** unless
anchored (per the dia2 README: "voices vary per generation … use with
prefix … for stable output"). To pin a voice, use the richer **`/tts`**
endpoint with `voice_mode: clone` + a `clone_reference_filename`.

The image bakes **43 curated voices** at `/app/voices` (singles +
`[S1]/[S2]` dialogue pairs like `Abigail_Taylor.wav`), but `/tts`'s clone
path only reads the **reference_audio** dir — so they're **staged** into it:

```bash
# one-time per host (writes through the shared /worktank/dia/reference_audio mount):
docker exec dia2-2b sh -c 'cp -n /app/voices/* /app/reference_audio/'
```

After staging, `GET /get_reference_files` lists them and asset-engine's
`clone_reference_filename` picker (sourced from that endpoint) offers a
stable, known voice. Restic-included, so it survives once staged.

## Notes

- **Model switching** within an instance is config.yaml-driven (the mounted
  `config.yaml` pins `model.repo_id`); the Web UI can hot-swap live but only
  the mounted config survives recreate.
- Endpoints: `/tts` (rich: cfg_scale/temperature/top_p/cfg_filter_top_k/
  voice_mode/clone — what asset-engine targets), `/v1/audio/speech`
  (OpenAI-compat; its `voice` param also resolves predefined voices by name),
  `/get_reference_files`, `/get_predefined_voices`, `/health`,
  `/api/model-status`, `/api/model-info`.