Organizations: Independent Researcher · University of Illinois at Urbana-Champaign · University of Florida · University of Minnesota - Twin Cities · IIT Roorkee
Sanskrit verse recitation is governed by Chandas, a strict prosodic framework defined by rhythmic Laghu (light) and Guru (heavy) syllables. Because existing Sanskrit speech datasets prioritize ASR or conversational TTS, they lack the metrical fidelity required for authentic verse synthesis. We introduce \textbf{ShlokaDhara}, a 43-hour, multi-speaker Sanskrit speech corpus capturing diverse meters, recitation styles, and melodic variations. We specifically focus on Text-to-Recite, a synthesis task explicitly requiring both prosodic accuracy and metrical correctness. Benchmarking state-of-the-art TTS architectures via an automated UTMOS (neural MOS) predictor demonstrates that XTTS-v2 achieves the highest perceptual quality. We further validate the XTTS-v2 generated samples using two paradigms: subjective assessments by 16 Sanskrit experts and novel empirical metrics, specifically, syllable-level Laghu-Guru duration analysis to objectively quantify structural metrical awareness. Ultimately, this work establishes a robust dataset and evaluation benchmark for prosody-aware computational Sanskrit recitation. Generated Samples: https://shlokadhara.vercel.app/
Figures & tables
Figure 1: Syllable-level Laghu (L, light) and Guru (G, heavy) annotation for the śloka Vidyā dadāti vinayaṃ, vinayād yāti pātratām (“Knowledge gives humility; from humility comes worthiness”).
Dataset
Language
Hrs
Speakers
Type
Vāksañcayaḥ
Sanskrit
78+
27
ASR
Kathbath
Multiple
191.8
Multiple
ASR
IndicVoices
Multiple
23K+
51K+
Conversational
Shrutilipi
Multiple
20
Multiple
Broadcast
Vedavani
Sanskrit
54
Limited
Vedic Chant
Rasa
Multiple
20/spk
44
Expressive TTS
Table 1: Comparison of existing Sanskrit and Indic datasets with ShlokaDhara .
Figure 2: ShlokaDhara Dataset Creation Pipeline
Metric
Value
Metric
Value
Duration (hours)
43.24
Avg. Duration (s)
8.5
Avg. Utterance Length
6.02 tokens
# Utterances
18,322
Avg. SNR (dB)
32.03
Spectral Centroid (Hz)
2185.3
Mean Pitch (Hz)
210.97
Pitch Std. Dev. (Hz)
42.22
Character Vocabulary
64
Word Vocabulary
18218
Table 2: Summary statistics of the ShlokaDhara dataset. All acoustic features are averaged over utterances.
Figure 3: Top sources other than Bhagavad Gita .
Figure 4: Semantic and acoustic PCA projections.
Figure 5: Average time-normalized F0 contours (Z-scores) of 13 speakers reciting Bhagavad Gita Chapter 2 . Highlighting three distinct profiles (highly dynamic, moderate, flat/monotonic) against the remaining speakers demonstrates the dataset’s broad melodic diversity.
Model
Seen
Unseen
UTMOS ↑
CMOS ↑
UTMOS ↑
CMOS ↑
VITS
3.12
-0.0515
2.78
-0.0611
NaturalSpeech-V2
3.52
-0.0503
2.82
-0.0606
EnGenTTS
3.87
-0.0492
3.01
-0.0598
Parler-TTS
3.66
-0.0442
2.56
-0.0555
Voxtral
3.52
-0.0411
3.12
-0.0562
Table 3: Results of TTS models fine-tuned on the proposed ShlokaDhara dataset (UTMOS range is 0 to 6)
Figure 6: Evaluation of Text-to-Recite Task
Evaluation Criteria
Score (mean ± 95% CI )
Clarity and naturalness of speech
2.91{\color[rgb]{0,0,0}\pm 0.14}
Accuracy of pronunciation
3.01{\color[rgb]{0,0,0}\pm 0.14}
Correctness of Sandhi and word formation
3.14{\color[rgb]{0,0,0}\pm 0.14}
Preservation of rhythmic and metrical structure
2.92{\color[rgb]{0,0,0}\pm 0.14}
Overall quality of recitation
2.92{\color[rgb]{0,0,0}\pm 0.13}
Table 4: Human evaluation results for the proposed Text-to-Recite task. (CI=confidence interval)
Provides hyperparameters and setup for reproducibility.
Appendix
Table A.1: Table of Contents of the Appendix Section
Speaker
Hours ↓
Speaker
Hours ↓
spk012
3.42
spk011
2.41
spk014
3.35
spk017
2.38
spk002
3.31
spk004
2.32
spk015
3.10
spk001
2.20
spk007
2.93
spk013
1.95
spk003
2.66
spk006
1.52
Appendix
Table A.2: Speaker-wise duration contribution in hours for complete dataset, sorted in descending order.
Figure A.1: Speaker-wise distribution for Bhagavad Gita recited by 12 speakers, with some chapters from one speaker
Source
Hours ↓
Speakers
Chandas
Gita Chapter 2
3.29
13
Anushtup, Upajati, Indravajra, Upendravajra
Gita Chapter 18
3.11
11
Anushtup, Trishtubh
Gita Chapter 11
3.08
12
Anushtup, Upajati, Indravajra, Upendravajra
Vishnusahastranam
2.54
6
Anushtup, Shalini, Vanshastha, Indravajra
Gita Chapter 6
2.08
12
Anushtup
Gita Chapter 4
2.04
13
Anushtup
Appendix
Table A.3: Source-wise duration distribution (in hours), number of unique speakers, and associated chandas for each source.
Figure A.2: Laghu-Guru Duration Analysis Plot
ID
# Laghu
# Guru
Laghu (s)
Guru (s)
Ratio
1
13
19
0.232
0.401
1.72
2
12
20
0.321
0.358
1.12
3
18
16
0.201
0.399
1.98
4
18
16
0.172
0.321
1.87
5
32
16
0.289
0.395
1.37
6
16
8
0.272
0.370
1.36
Appendix
Table A.4: Verse-by-verse breakdown of syllable counts and empirical durations. ID: Sample ID, # Laghu / # Guru: Laghu/Guru Count, Laghu / Guru (s): Average Laghu/Guru Duration and Guru to Laghu Ratio. For averaging the count and duration, arithmetic mean has been used. For averaging the ratio (last column), geometric mean has been used.
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.
Poetry generation in Sanskrit typically requires the verse to be semantically coherent and adhere to strict prosodic rules. In Sanskrit prosody, every line of a verse is typically a fixed length sequence of syllables adhering to prescribed binary patterns of syllable weights. We observe that instead of treating a verse as a monolithic sequence, segmenting them as grouped-lines leads to significant improvement in semantic coherence by 10% with comparable metrical adherence. Specifically, Pingala, our proposed decoding approach is designed to encourage every line to have well-formed words and our token selection biases the model towards it by preferring longer tokens. Writing in Sanskrit follows phonemic orthography, hence using a phonetically aware transliteration scheme, SLP1, increased the metrical alignment by 46% with comparable semantic similarity, for a instruction fine-tuned large language models like Phi-4. We also introduce a new approach for reference-free evaluation using cross-encoders which achieved better alignment with true poetry instances.
Zero-shot text-to-speech (TTS) typically conditions on a short speech prompt, which conveys timbre but provides only limited evidence of a speaker's prosodic patterns. We formulate Long-Prompt Zero-Shot TTS, a task that uses long speech prompts at inference time to reproduce speaker-specific prosody without target-speaker fine-tuning. We introduce LP-TTS-Bench, a bilingual benchmark of 37 speakers with 63.95 prompt hours and 2,728 held-out test utterances. Beyond intelligibility and timbre similarity, LP-TTS-Bench evaluates timing, prominence, intonation, and model-based prosody metrics, enabling multidimensional evaluation of prosody reproduction. As a reference system, we develop LPTTS, whose hierarchical Long-Prompt Encoder compresses long speech prompts into a fixed number of continuous prosody vectors. Experiments show improvements in several prosodic dimensions over a short-prompt baseline, with overall gains as prompt duration increases up to approximately 10 minutes and saturation at longer durations.
Dekun Chen, Renjie Xu, Yingda Shen +2
The Chinese University of Hong Kong, Shenzhen · Amphion Technology Co., Ltd.