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. Demo: https://charlesnii.github.io/Emilia2-demo/.
Figures & tables
Resource
Hours
Language
Long
Dialogue
Instruction
NV
Emilia ( He et al., 2025 )
216,000
6
✗
✗
✗
✗
MinutesSpeech ( Nishimura et al., 2025 )
40,000
1
✓
✗
✗
✗
OpenDialog ( Zhu et al., 2026a )
6,800
2
✗
✓
✗
✗
DNASpeech ( Cheng et al., 2025 )
18
1
✗
✓ †
✓
✗
ParaSpeechCaps ( Diwan et al., 2025 )
2,800
1
✗
✗
✓
✗
CapSpeech ( Wang et al., 2026b )
33,600
1
✗
✗
✓
✓
Table 1: Emilia2 versus existing TTS training datasets. Long denotes duration longer than 30 s. Dialogue denotes multi-speaker dialogue speech. Instruction denotes natural language instructions. NV denotes nonverbal vocalizations. † Dialogue context prompts for utterance synthesis.
Figure 1: Pipeline overview of Emilia2-Pipe . In-the-wild recordings are enhanced, diarized, transcribed, and merged into single-speaker or multi-speaker audio. Segments are annotated with natural language style instructions and NV tags, then reassembled with speaker and timing information into single-speaker long-form speech and multi-speaker dialogue examples.
Figure 3
View
Speech data
Instruction
NV
Examples
Audio (h)
Examples
Audio (h)
Examples
Audio (h)
Emilia2-Short
99,841,007
146,325
3,994,566
5,102
474,078
854
Emilia2-Long
4,263,913
155,688
651,060
15,317
353,929
9,572
Emilia2-Dialogue
1,538,239
96,452
120,815
4,633
104,999
4,200
Total
105,643,159
398,467
4,766,441
25,053
933,006
14,626
Table 2: Scale and annotation coverage of Emilia2 . The three views share source recordings and the Total row sums view entries. NV counts examples containing NV tags.
Language
Cases
Single-speaker
Dialogue
Audio (h)
Avg. (min)
Range (min)
Chinese
59
18
41
11.6
11.84
0.9–29.3
English
59
24
35
11.7
11.86
1.0–30.0
Total
118
42
76
23.3
11.85
0.9–30.0
Table 3: Composition of Emilia2 -testset. Hours cover target recordings only. Avg. and Range are the average and range of target recording durations.
Model
Params.
Chinese
English
ACC ↑
SIM ↑
CER ↓
ACC ↑
SIM ↑
WER ↓
GT
–
99.33
0.88
3.00
98.38
0.85
3.00
Higgs Audio v2
3B
75.00
0.74
75.46
72.84
0.67
81.38
VibeVoice-1.5B
1.5B
93.78
0.81
4.61
95.98
0.75
5.66
MOSS-TTSD
8B
99.28
0.87
2.18
99.22
0.81
5.83
Emilia2-AR
0.6B
98.57
0.84
3.67
95.05
0.74
6.82
Table 4: Objective results on Emilia2 -testset. ACC is accuracy over multi-speaker cases. Single-speaker cases remain in SIM and CER/WER. Bold and underlining mark the best and second-best per column.
Model
TTSD-Eval
Long-TTS-Eval
Chinese
English
Chinese
English
ACC ↑
SIM ↑
CER ↓
ACC ↑
SIM ↑
WER ↓
CER ↓
SIM ↑
WER ↓
SIM ↑
Higgs Audio v2
60.97
0.63
79.32
70.60
0.62
47.54
73.65
0.60
73.30
0.47
VibeVoice-1.5B
88.25
0.73
9.21
93.58
0.68
10.18
16.98
0.67
11.20
0.59
MOSS-TTSD
95.36
0.80
6.08
95.00
0.74
11.09
8.22
0.74
8.54
0.63
Emilia2-AR
92.89
0.75
6.92
92.66
0.65
12.25
14.42
0.75
17.76
0.67
Table 5: Long-form synthesis on TTSD-Eval and Long-TTS-Eval. ACC denotes speaker-attribution accuracy, SIM speaker similarity, and CER/WER character/word error rate. Bold and underlining mark the best and second-best scores per column.
Short-form
Chinese
English
Model
APS
DSD
RP
Avg.
APS
DSD
RP
Avg.
VoxInstruct
47.5
52.3
42.6
47.5
54.9
57.0
39.3
50.4
FlexiVoice-Base
56.7
59.1
59.5
58.4
63.6
75.0
60.6
66.4
Qwen3-TTS-12Hz-VD
85.2
81.1
65.1
77.1
82.9
82.4
68.4
77.9
Emilia2-AR
82.0
72.5
74.0
76.2
83.0
83.5
68.5
78.3
Emilia2-NAR
69.2
59.8
55.6
61.6
67.4
66.5
49.8
61.2
Table 6: Instruction following at two timescales. Short-form results are on InstructTTSEval, and long-form results on instruction-conditioned Emilia2 -testset cases. In the short-form block, bold and underlining mark the best and second-best scores per column.
Short-form
Single-label subset
Multi-label subset
Model
Chinese
English
Chinese
English
PCER ↓
CER ↓
SIM ↑
PWER ↓
WER ↓
SIM ↑
PCER ↓
CER ↓
SIM ↑
PWER ↓
WER ↓
SIM ↑
Orpheus-TTS
88.77
11.36
–
71.92
9.03
–
84.85
19.83
–
71.46
8.68
–
SMIIP-NV-CV2
75.64
8.80
0.72
56.80
17.92
0.58
77.20
10.66
0.72
54.49
20.93
0.58
CosyVoice 3
57.69
3.85
0.76
62.75
7.87
0.70
61.94
4.75
0.72
57.84
6.39
0.72
Emilia2-AR
39.23
5.39
0.67
55.59
10.96
0.62
49.48
6.25
0.67
57.19
9.10
0.62
Table 7: NV generation at two timescales. Short-form results are on NV-Bench, and long-form results on NV-conditioned Emilia2 -testset cases. In the short-form block, bold and underlining mark the best and second-best scores per column.
Figure 4: Stratified Chinese CER on Emilia2 -testset. Top: by speaker count, bottom: by target duration.
Appendix figures & tables21 assets
Supplementary material from the paper’s appendix.
Appendix
Benchmark
Coverage
FireRedLID
Baseline
Baseline Acc.
FLEURS test
82 langs.
97.18
SpeechBrain
92.91
CommonVoice test
74 langs.
92.07
Whisper
80.81
Chinese dialects
10+ dialects
88.47
Dolphin
69.01
Appendix
Table 8: Published FireRedLID accuracy (%). Results reported by Xu et al. [2026] .
Language
SpeechBrain Acc. (kept)
FireRedLID Acc. (kept)
Δ Acc.
Bengali (bn)
83.3 (54)
92.1 (63)
+8.8
German (de)
89.5 (76)
87.3 (63)
−2.2
English (en)
98.9 (93)
100.0 (95)
+1.1
Spanish (es)
71.7 (53)
91.4 (70)
+19.7
French (fr)
88.9 (81)
90.3 (72)
+1.4
Japanese (ja)
87.7 (73)
95.3 (86)
+7.6
Appendix
Table 9: In-house nine-language LID audit at confidence τ≥0.8 . Accuracy is computed on retained predictions. Parentheses give retained item counts from the same evaluation pool. Bold marks the higher per-language accuracy, not greater coverage.
Benchmark
Coverage
Qwen3-ASR-1.7B
Whisper-L3
FLEURS-en
English
3.35
4.08
FLEURS-zh
Chinese
2.41
4.09
MLS
8 langs.
8.55
8.62
CommonVoice
13 langs.
9.18
10.77
FLEURS
12 langs.
4.90
5.27
FLEURS extended
30 langs.
12.60
8.16
Appendix
Table 10: Published Qwen3-ASR-1.7B error rates (%). Results reported by Shi et al. [2026] .
Method
Acc.
RTF
WavLM
0.938
0.0028
wav2vec2
0.936
0.0015
Qwen2-Audio
0.919
0.0077
Appendix
Table 11: Gender classification on FLEURS (9 languages). Bold marks the deployed classifier. Higher Acc. and lower RTF are better. RTF is processing time divided by audio duration.
Method
Acc.
UAR
FireRedLID → Voxlect
0.757
0.743
Voxlect (Whisper-L3)
0.743
0.733
FireRedLID
0.331
0.243
Appendix
Table 12: Mandarin-accent classification on KeSpeech (7 classes). Bold marks the deployed method. Higher Acc. and UAR are better.
Measure
Acc.
UAR
Characters/s
0.894
0.937
Phones/s
0.732
0.830
Words/s
0.732
0.819
Appendix
Table 13: Speaking-rate measures on PSC ( n=246 ). Bold marks the deployed measure. Higher Acc. and UAR are better.
Set
Lang.
n
No gate
τ=0.7
WA
UA
WF1
Cov.
WA
UA
WF1
MSP-Podcast
en
3,256
0.49
0.45
0.45
28%
0.72
0.54
0.66
MELD
en
2,608
0.53
0.39
0.53
20%
0.71
0.52
0.71
M3ED
zh
4,194
0.53
0.42
0.54
25%
0.73
0.59
0.73
BIIC-Podcast
zh
10,149
0.57
0.37
0.56
29%
0.79
0.55
0.77
Appendix
Table 14: Seven-class emotion recognition with and without confidence gating. Coverage is the fraction retained at τ=0.7 . Gated scores are computed on retained clips.
Figure 5: Topical composition of the source data. Slice labels report the share of data hours. Other (38.5%) combines 25 lower-frequency topics.
Duration
Emilia2-Long
Emilia2-Dialogue
Examples
(%)
Hours
(%)
Examples
(%)
Hours
(%)
30–60 s
2,337,079
54.81
27,015
17.35
557,474
36.24
6,696
6.94
1–5 min
1,571,270
36.85
50,868
32.67
737,290
47.93
25,321
26.25
5–10 min
196,474
4.61
22,972
14.76
114,598
7.45
13,441
13.94
10–15 min
73,138
1.72
14,827
9.52
47,164
3.07
9,589
9.94
15–20 min
33,896
0.79
9,722
6.24
25,552
1.66
7,363
7.63
Appendix
Table 15: Example-duration distributions of Emilia2-Long and Emilia2-Dialogue . Each view is counted separately. Hours and percentages are rounded, so their sums may differ slightly from the totals.
Language
Emilia2-Short
Emilia2-Long
Emilia2-Dialogue
Hours
Share (%)
Hours
Share (%)
Hours
Share (%)
English
36,808
25.16
45,311
29.10
36,238
37.57
Chinese
36,739
25.11
54,257
34.85
17,383
18.02
German
24,646
16.84
19,703
12.66
21,701
22.50
Japanese
14,940
10.21
8,491
5.45
7,710
7.99
Arabic
8,400
5.74
7,928
5.09
3,253
3.37
Appendix
Table 16: Language distribution of the three training views. Each example is assigned to its dominant language; Other aggregates the remaining languages. Long and Dialogue hours are uniformly normalized to the corresponding totals in Table 2 . Hours and shares are rounded.
Figure 6: Expressive annotation distributions of Emilia2 . (a–c) Predicted attributes over 5,096,107 voice records from 4,766,441 instruction-annotated examples, with one record per speaker in Dialogue. Attribute distributions exclude missing labels. In (b), age forms the inner ring and voice-group labels the outer ring. (d) NV tag occurrences in 933,006 NV-positive records, grouping Surprise and Question variants. Records and occurrences may overlap across views.
Benchmark
Type
Clips
Avg. dur.
Speaker
Open source
Prompt audio
MinutesSpeech-test [ Nishimura et al., 2025 ]
Long
1,221
0.84 min
1
✗
✓
LibriSpeech-Long [ Park et al., 2024b ]
Long
960
3.19 min
1
✓
✗
Long-TTS-Eval [ Wang et al., 2026a ]
Long
694
13.4 min †
1
✓
✗
MultiDialog [ Park et al., 2024a ]
Dialogue
831
2.46 min
2
✓
✗
ZipVoice-Dialog testset [ Zhu et al., 2026a ]
Dialogue
637
0.38 min
2
✓
✓
TTSD-Eval [ Zhang et al., 2026b ]
Both
100
10.0 min †
2
✓
✓
Appendix
Table 17: Comparison of long-form TTS testsets. Average duration is per target recording, using published or ground-truth audio durations when available. † denotes an estimate from target text at 3 characters per second. “Prompt audio” denotes a reference recording for each speaker.
English speakers
Chinese speakers
Target duration
1
2
3
4
1
2
3
4
Total
< 5 min
8
4
2
1
5
6
4
2
32
5–10 min
4
2
4
2
3
4
3
0
22
10–15 min
3
4
4
1
2
4
2
1
21
15–20 min
3
4
1
0
3
2
2
3
18
20–25 min
3
2
1
0
3
4
2
0
15
Appendix
Table 18: Emilia2 -testset cases by duration and speaker count. Each language contains 59 recordings.
Figure 7: English WER, prompt SIM, and speaker ACC versus target duration. Columns show one through four speakers. Curves compare matched GT, Emilia2-AR, and Emilia2-NAR recordings. Single-speaker ACC is undefined. Cell counts are in Table 18 .
Figure 8: Chinese CER, prompt SIM, and speaker ACC versus target duration. Systems and speaker-count groups follow Figure 7 . In the four-speaker column, markers occur only at scored bins. Cell counts are in Table 18 .
Figure 9: Emilia2 model architectures. (a) Emilia2-AR combines autoregressive token prediction with flow-matching reconstruction. Generated speaker IDs select the associated reference voice. (b) Emilia2-NAR learns masked acoustic-token prediction and generates speech through iterative parallel unmasking.
Stage
Max. duration
Learning rate
Max. token length
Short-form
30 s
2×10−4
32,768
Long-form
30 min
2×10−4
32,768
Dialogue
30 min
2×10−4
50,480
Expressive SFT
5 min
1×10−5
50,480
Appendix
Table 19: Representative Emilia2-AR training settings.
Model
Inference
Chinese
English
Overall N-MOS ↑
CER ↓
N-MOS ↑
WER ↓
N-MOS ↑
Ground truth
Recorded
–
4.55 ± 0.31
–
4.55 ± 0.27
4.55 ± 0.20
Emilia2-AR
Chunked
4.36
3.30 ± 0.16
2.44
4.08 ± 0.28
3.69 ± 0.24
Whole sequence
3.67
4.05 ± 0.23
6.82
4.10 ± 0.24
4.08 ± 0.16
Emilia2-NAR
Chunked
1.73
3.90 ± 0.28
1.52
4.00 ± 0.21
3.95 ± 0.17
Whole sequence
8.62
4.05 ± 0.31
6.62
4.10 ± 0.38
4.08 ± 0.24
Appendix
Table 20: Chunked versus whole-sequence inference on Emilia2 -testset. N-MOS is reported as the case mean ± 95% confidence interval. Bold marks the better result within each model and language.
Model
English
Chinese
SIM-o ↑
WER ↓
SIM-o ↑
CER ↓
Ground-truth
0.73
2.14
0.76
1.25
Higgs Audio v2
0.65
2.60
0.72
2.62
VibeVoice-1.5B
0.69
3.04
0.74
1.16
MOSS-TTSD
0.68
9.83
0.76
5.84
Emilia2-AR
0.61
1.85
0.69
1.46
Appendix
Table 21: Short-form results on Seed-TTS-Eval. Bold marks the best available generated-speech score in each column.
Model
WER ↓
SIM-o ↑
Higgs Audio v2
49.51
0.70
VibeVoice-1.5B
85.41
0.75
MOSS-TTSD
40.25
0.80
Emilia2-AR
7.67
0.71
Emilia2-NAR
10.34
0.79
Appendix
Table 22: Macro-averaged results on MiniMax-Multilingual-24. Languages are weighted equally. WER is in percent. Bold marks the best score per metric.
Language
Emilia2-AR
Emilia2-NAR
WER ↓
SIM-o ↑
WER ↓
SIM-o ↑
Arabic
7.34
0.71
13.79
0.76
Cantonese
37.15
0.75
28.21
0.78
Chinese
3.72
0.73
1.25
0.79
Czech
11.97
0.70
18.90
0.76
Dutch
4.47
0.69
6.91
0.72
Appendix
Table 23: Per-language Emilia2 results on MiniMax-Multilingual-24. WER is in percent recomputed from the 24 displayed language rows. Bold marks the best value for each metric in each row.
Neural Text-to-Speech (TTS) systems achieve remarkable quality on short utterances but long-form speech generation shows prosodic drift, speaker inconsistencies and sentence boundary artifacts. Existing approaches either compress sequences, increase context length or naively concatenate independently synthesized chunks. We present an inference-time approach called MagpieTTS-LF that enables MagpieTTS to produce coherent long-form speech without model retraining. Our method introduces three key innovations: (1) soft attention priors to guide monotonic alignment while preserving past and future context; (2) a stateful inference algorithm that maintains context across sentence chunks, ensuring prosodic continuity; (3) history-aware text encoding that uses past text for discourse-level prosodic planning. Experiments on long texts show significant improvements in long-range intelligibility, prosodic coherence, speaker consistency, and boundary naturalness compared to other baselines.
Recent full-stream text-to-speech (TTS) models enable low-latency personalized synthesis as input text arrives. However, sustaining generation beyond the lengths seen during training without additional training remains largely unexplored. We systematically analyze VoXtream2, which extends generation by periodically discarding past context while retaining the acoustic prompt. We find that this strategy does not prevent degradation: pronunciation accuracy deteriorates near the end of the training range due to the range of positions in which the model operates rather than the loss of past context, while speaker similarity gradually declines throughout generation. Based on these findings, we propose two training-free modifications that reduce the attention window and strengthen attention to the acoustic prompt. Experiments show that these modifications preserve pronunciation accuracy and speaker similarity throughout long-form generation while retaining the naturalness of the original model. Ablations further show complementary effects of the two modifications.
Jeseung Lim, Jiseong Woo, Heeseung Kim
Department of Physics, University of Seoul, Seoul, Republic of Korea · Department of AI, University of Seoul, Seoul, Republic of Korea
Audiobook narration, conversational agents, and audiovisual dubbing require fine-grained control over emotion and pacing within an utterance, yet many text-to-speech (TTS) systems remain limited to utterance-level conditioning. We propose a unified post-training framework for natural-language segment-level emotion and duration control. Supervised fine-tuning (SFT) establishes instruction following, while group relative policy optimization (GRPO) further improves control accuracy using emotion, duration, content, and speaker rewards. Without modifying the pretrained architecture or adding inference-time control modules, our method improves fine-grained controllability while preserving intelligibility and speaker identity, demonstrating post-training as an effective way to extend pretrained TTS models.
Lianru Gao, Yujie Guo, Yong Qin
TMCC, College of Computer Science, Nankai University, Tianjin, China