Text-to-speech (TTS) is a "one-to-many" problem: one text can be spoken in many valid ways, so its realizations form a distribution whose spread depends on the input. Most TTS models do not model this spread explicitly: they leave it implicit in the sampling process or fix it in advance. To model this distribution explicitly, we propose BELLE (Bayesian evidential learning with language modelling), which places a Bayesian prior over the mean and variance of each mel-spectrogram frame and predicts this prior from the input, while keeping the architecture and single-pass sampling of common continuous-valued autoregressive TTS models such as MELLE. Estimating this variance requires several recordings of each text, whereas standard TTS corpora provide one. We therefore construct two parallel-speech corpora: ParallelVox from real human recordings and LS+6TTS, which adds speech from six pre-trained TTS teachers to LibriSpeech (LS). Trained on these corpora, BELLE outperforms MELLE under the same training conditions, is competitive with larger TTS systems on the objective metrics of LibriSpeech test-clean, and supports streaming generation. Analyses of repeated human readings show that real speech has the properties the Bayesian prior assumes and that the variance predicted by BELLE follows the variation across these readings. Parallel speech improves BELLE, and synthetic parallel speech performs comparably to real parallel speech at the same volume and is far easier to scale.
Figures & tables
Figure 1: Overview of BELLE. Left: the mel-based AR TTS architecture (Sec. 3.2 ). Right: the Sampling Module, which predicts a Bayesian prior over the mean and variance of each frame and samples the frame from it (Sec. 4 ).
Figure 2: Interleaved text and mel-spectrogram chunks in BELLE-stream.
System
Training data
Continuation
Cross-Sentence
Diversity
Subjective
WER-H ↓
SIM-o ↑
WER-H ↓
SIM-o ↑
Cos. ↑
MOS ↑
SMOS ↑
Ground Truth
-
2.15
0.668
2.15
0.672
46.7
4.20
3.81
VALL-E
ParallelVox
7.46
0.340
16.3
0.452
17.4
-
-
MELLE
ParallelVox
3.06
0.406
5.60
0.538
19.9
-
-
BELLE
ParallelVox
3.00
0.449
5.73
0.570
21.5
-
-
BELLE
LS+6TTS-881h
2.57
0.510
4.05
0.634
20.3
-
-
Table 1: Controlled comparison under a shared training and evaluation pipeline. WER-H and SIM-o are measured on LibriSpeech test-clean, and Cos. ( ×10−4 ) on the ParallelVox single-speaker repeat test set. MOS and SMOS were collected only for Ground Truth and the models trained on LS+6TTS.
System
WER-H ↓
SIM-o ↑
SMLLE
6.37
0.489
IST-LM
4.53
0.653
BELLE-stream
4.01
0.557
Table 3: Cross-Sentence results of streaming systems: SMLLE ( Sun et al., 2025 ) , IST-LM ( Yang et al., 2024 ) , and BELLE-stream trained on LS+6TTS.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Source
Recordings
Hours
Speakers
License
LibriHeavy ( Kang et al., 2024 )
312,959
1,017.7
3,392
CC BY 4.0
Parallel Audiobook Corpus
84,791
134.4
53
Public Domain
VCTK 0.92 ( Yamagishi et al., 2019 )
34,981
20.7
109
CC BY 4.0
CMU ARCTIC ( Kominek and Black, 2004 ) + L2-Arctic ( Zhao et al., 2018 )
18,625
15.5
42
Mixed
Common Voice 17.0 ( Ardila et al., 2020 ) , parallel subset
284
0.2
255
CC-0
Total
451,640
1,188.5
3,851
Appendix
Table 4: Source corpora of ParallelVox after VAD trimming and before WER filtering.
WER ≤
MOS ≥
Texts
Recordings
Hours
10%
—
123,640
344,303
903.0
10%
3.8
60,012
175,356
400.1
10%
4.0
38,618
114,741
244.4
5%
—
92,734
262,337
638.5
3%
—
73,683
214,831
474.9
Appendix
Table 5: Size of ParallelVox under different WER and MOS thresholds. The first row is the setting used in this paper.
Human
Each teacher
Continuation
Cross-Sentence
WER-H ↓
SIM-o ↑
WER-H ↓
SIM-o ↑
0.16
0.14
2.64
0.415
5.17
0.453
0.22
0.13
2.57
0.409
4.73
0.449
0.28
0.12
2.78
0.406
5.47
0.444
0.34
0.11
2.76
0.405
5.47
0.446
Appendix
Table 6: Weights of the human recording and of each teacher realization in Eq. ( 13 ), on LS-100+6TTS . The second row is the setting used in this paper.
Figure 3: Instructions shown to the raters in the subjective evaluation.
Group source
Duration
Median F0
F0 s.d.
Energy
Energy s.d.
Ground Truth (human readings)
0.54
0.70
0.61
0.21
0.65
BELLE ParallelVox
0.07
0.37
0.25
0.34
0.33
BELLE LS+6TTS
0.19
0.38
0.25
0.29
0.29
Appendix
Table 7: Spearman’s ρ between the within-group variation of prosodic quantities and the within-group WavLM distance.
Metric
ParallelVox
LS
LS+6TTS
Continuation WER-H
−0.07 ( p=0.52 )
−0.39 ( p<0.001 )
−0.46 ( p<0.001 )
Continuation SIM-o
+0.044 ( p<0.001 )
+0.005 ( p=0.004 )
+0.031 ( p<0.001 )
Cross-Sentence WER-H
+0.13 ( p=0.56 )
−3.75 ( p<0.001 )
−0.84 ( p<0.001 )
Cross-Sentence SIM-o
+0.032 ( p<0.001 )
+0.015 ( p<0.001 )
+0.036 ( p<0.001 )
Diversity (Cos.)
+1.67 ( p=0.007 )
+6.90 ( p<0.001 )
+2.60 ( p<0.001 )
MOS
–
–
+0.19 ( p=0.021 )
Appendix
Table 8: BELLE − MELLE differences and p -values, from a paired bootstrap for the objective metrics and a paired Wilcoxon test for MOS and SMOS. MOS and SMOS were collected only on LS+6TTS.
Sampling
Flux loss
Continuation
Cross-Sentence
WER-C
WER-H
SIM-r
SIM-o
WER-C
WER-H
SIM-r
SIM-o
✗
✓
4.30
5.00
0.382
0.344
12.04
13.07
0.392
0.351
✓
✗
2.50
3.25
0.388
0.347
5.61
6.37
0.417
0.372
✓
✓
2.04
2.66
0.398
0.362
4.12
4.84
0.433
0.395
Appendix
Table 9: Ablation of the Sampling Module and the flux loss on BELLE LS-100+2TTS . ✓ indicates enabled, and ✗ indicates disabled.
Sampling Module and loss
Continuation WER-H ↓
Cross-Sentence WER-H ↓
None
5.00
13.07
Gaussian, KL toward unit variance (MELLE)
3.24
6.83
Gaussian, NLL with free variance
3.88
7.31
NIG, evidential loss (BELLE)
2.66
4.84
Appendix
Table 10: Sampling Modules and their losses on LS-100+2TTS .
We distill Kokoro-82M, a widely used open text-to-speech model with 54 voices, into Paradee, an 8.07M-parameter model that speaks one of them. Paradee keeps Kokoro's architecture with much narrower layers, and each of its two halves is trained separately against the frozen teacher. It has 10x fewer parameters and needs 15x less compute. We first synthesize a corpus with the teacher and keep its durations, pitch, energy and phoneme features. We then train a small text side to predict these values, and a small decoder to turn the teacher's saved values into the teacher's audio, first with spectral losses and then adversarially. Finally, we connect the two halves and quantize the weights to int8. It needs no alignment learning and no joint training, and it runs on one laptop. Stored in int8, Paradee is 8.5 MB, runs 25x faster than real time on one CPU thread, and scores 4.41 on UTMOS against the teacher's 4.52. The student initially kept a slight buzz, which we trace to the phase of voiced speech between 2 and 8 kHz. A phase-locking filter applied after synthesis removes most of it, with no training and no extra parameters. Code, model files and audio samples are at https://github.com/sahilmahendrakar/paradee
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
Bangla, the seventh most spoken language in the world, remains under-resourced for neural text-to-speech. Public Bangla speech corpora are dominated by short read-prompt utterances collected for speech recognition, leaving long-form prosody and consistent single-speaker narration uncovered. We present BanglaKontho, a single-speaker Bangla TTS corpus of 20 hours derived from professional audiobook recordings: 7,050 segmented utterances with verified transcripts at 24 kHz. We also release a reusable Bangla text normalizer covering Bangladeshi-style digit grouping, currency and date expressions, Danda punctuation and Unicode normalization, together with the full preprocessing pipeline. An MB-iSTFT-VITS baseline trained from scratch reaches 9.5% WER and 4.46 naturalness MOS, against 16.0% and 3.16 for the same architecture retrained on the 12-hour IndicTTS-Bn corpus. The corpus is released openly under CC BY-NC 4.0.