Skip to content
Bud's Wiki

Dia2 streaming TTS release — 1B/2B Apache-2.0 model with prefix conditioning

TL;DR

Section titled “TL;DR”

A short third-party announcement tweet from @Tu7uruu noting the release of Dia2, a streaming text-to-speech model that generates audio incrementally as text arrives. The tweet lists four headline features — on-the-fly audio generation without full text input, real-time conversational streaming, prefix conditioning for contextual responses, and 1B / 2B sizes — and states the release is fully open-source under Apache 2.0. No paper, repo URL, or model card is linked from the tweet itself; this page captures the announcement as it circulated.

Key claims

Section titled “Key claims”
  • Dia2 generates audio on the fly without needing the full text input in advance [tweet body].
  • The model supports real-time conversational streaming, pitched at agents and AI voice applications [tweet body].
  • Prefix conditioning is offered as the mechanism for smoother, contextual voice responses [tweet body].
  • Two model sizes ship at release: 1B and 2B parameters [tweet body].
  • Licensing is Apache 2.0 (fully open-source) [tweet body].

Method

Section titled “Method”

Not described in the tweet. The four feature bullets imply (but do not specify) a causal/streaming language-model-style decoder over audio tokens with the ability to condition on a prefix of prior audio + text context — the same family of designs as recent open-source streaming TTS systems. There is no attached arxiv link, GitHub repo, or HuggingFace URL in the tweet, so the architecture details, training data, vocoder choice, latency numbers, and benchmark results are all unstated here.

Results

Section titled “Results”

None reported in the tweet. The post is an announcement, not a benchmark.

Why it’s interesting

Section titled “Why it’s interesting”

This sits alongside a cluster of recent open-weight streaming-TTS releases already on the wiki: CosyVoice 2: Scalable Streaming Speech Synthesis with Large Language Models is the canonical reference for a streaming zero-shot TTS architecture (Qwen2.5-0.5B backbone, FSQ tokenizer, N:M text-speech interleaving for chunked streaming), and Chatterbox Turbo — 350M Zero-shot TTS (Resemble AI) is a smaller 350M zero-shot competitor. Dia2’s “prefix conditioning” phrasing is consistent with the kind of audio-token-prefix scheme CosyVoice 2 uses to interleave text and speech tokens at low latency. Until a paper or repo surfaces, this is a pointer, not evidence: file as-is and append if a tech report appears.

See also

Section titled “See also”