arXiv Preprint

Emilia2: A Large-Scale Dataset for Controllable Long-Form Speech Generation

Qinke Ni1 Dekun Chen1 Junwen Qiu1 Huan Liao1 Xu Li1 Xingwen Han1 Wan Lin1 Kunyu Feng1 Yuxiang Wang1 Yusheng Mu1 Chaoren Wang1 Yingda Shen1 Mingjie Chen1 Haolong Zheng2 Xulin Fan2 Mark A. Hasegawa-Johnson2 Zhizheng Wu1

1The Chinese University of Hong Kong, Shenzhen    2University of Illinois at Urbana-Champaign

Code Testset Audio Samples
229Khours of speech
41languages
4.77Mexamples with style instructions
933Kexamples with nonverbal tags
30 minlongest testset recordings, up to 4 speakers
Emilia2-Dialogue example · 4 speakers · 30 min, one continuous recording 0:00 / 30:00
Figure 1. A long-form dialogue drawn as a score (illustrative). Each staff is one speaker and the waveform marks who holds the floor. Italic markings are natural-language style instructions, and bracketed tags are nonverbal vocalizations (NV). Hover to scrub.

Abstract

Generating long-form speech requires models to keep content, speaker identity, turn-taking, and expressive style consistent over time. Yet most open training corpora contain examples lasting only tens of seconds and open long-form benchmarks cover at most two speakers per recording. We introduce Emilia2, a 229K-hour multilingual dataset for controllable long-form speech generation. From in-the-wild data, Emilia2 constructs short utterances, single-speaker long-form speech, and multi-speaker dialogues, preserving transcripts, speaker identities, pauses, and acoustic transitions. Examples longer than five minutes account for 50.0% of single-speaker long-form hours and 66.8% of dialogue hours. To support expressive control, Emilia2 provides natural language instructions for 4.77M examples and nonverbal-vocalization tags for 933K examples. We also introduce Emilia2-testset, 23 hours of bilingual recordings of up to 30 minutes and four speakers, each speaker paired with reference audio. We train autoregressive (AR) and non-autoregressive (NAR) models on Emilia2. Our 0.6B-backbone NAR model achieves the lowest Chinese CER and English WER on TTSD-Eval (5.27%/9.93%) among compared systems, while our AR model is on par with Qwen3-TTS-12Hz-VD on InstructTTSEval (zh/en: 76.2/78.3 vs. 77.1/77.9). These results validate Emilia2 as a comprehensive data foundation for long-form speech generation and minute-scale expressive control.

Emilia2 Dataset

Emilia2-Pipe turns in-the-wild recordings into three views that share the same source audio. Long-form views keep the continuous recording, including native pauses and transitions between speakers, instead of concatenating trimmed clips. Annotated subsets add natural-language style instructions and NV tags placed among the spoken words.

Emilia2-Pipe overview: data construction (speech enhancement, diarization, transcription, merge), additional annotation (instruction and NV annotation), and the three resulting example types.
Figure 2. Overview of Emilia2-Pipe. Recordings are enhanced, diarized, transcribed, and merged; segments are annotated with style instructions and NV tags, then reassembled with speaker and timing information into long-form speech and dialogue examples.
Speaker 1 Speaker 2 Extracted example

Figure 3. Three views cut from one illustrative 12-minute source recording, with the size of each view in Emilia2. After deduplication, the corpus holds 229,388 hours across 41 languages.

Emilia2-testset. 118 human-verified Chinese and English recordings (23.3 h): 42 single-speaker long-form cases and 76 dialogues with 2–4 speakers, up to 30 min each. Every case provides a speaker-attributed script and reference audio for each speaker, and the whole script must be synthesized as a single recording. The testset will be released on Hugging Face (coming soon).

Explore the Data

Examples from each view of Emilia2, drawn from the training set with test-set recordings excluded. Long-form and dialogue examples are continuous recordings that keep native pauses and turn-taking; dialogue transcripts are speaker-attributed.

Browse the full dataset on 🤗 Hugging Face

Results

We train two models with a 0.6B backbone on Emilia2: Emilia2-AR (autoregressive with flow matching) and Emilia2-NAR (masked generative). Both are competitive with much larger systems on long-form dialogue and learn instruction and NV control, which no open-source baseline supports for long-form speech.

5.27% / 9.93%Chinese CER / English WER on TTSD-Eval with Emilia2-NAR, the lowest among compared systems
0.6Bbackbone parameters for both models, against 1.5B–8B for the dialogue baselines
76.2 / 78.3zh / en average on InstructTTSEval with Emilia2-AR, on par with Qwen3-TTS-12Hz-VD (77.1 / 77.9)

Bold and underlining mark the best and second-best generated-speech scores per column. Parameter counts cover the sequence backbone only.

Audio Samples

Speech generated by Emilia2-AR and Emilia2-NAR on Emilia2-testset. The long-form cases compare whole-sequence inference, which generates the complete script in one pass, with chunked inference, which generates short chunks of about 25–30 s independently and concatenates them. The instruction and NV cases come from the expressive-control evaluation sets built from Emilia2-testset cases shorter than 5 min.

BibTeX

If Emilia2 is useful in your work, please cite the paper.

emilia2.bib 
@article{ni2026emilia2,
  title   = {Emilia2: A Large-Scale Dataset for Controllable Long-Form Speech Generation},
  author  = {Ni, Qinke and Chen, Dekun and Qiu, Junwen and Liao, Huan and
             Li, Xu and Han, Xingwen and Lin, Wan and Feng, Kunyu and
             Wang, Yuxiang and Mu, Yusheng and Wang, Chaoren and Shen, Yingda and
             Chen, Mingjie and Zheng, Haolong and Fan, Xulin and
             Hasegawa-Johnson, Mark A. and Wu, Zhizheng},
  journal = {arXiv preprint},
  year    = {2026}
}