LREC 2026 · 2026 · Conference

Deepa P. Gopinath et al. (Kerala Development and Innovation Strategic Council / University of Hamburg) · → Paper · Demo: ? · Code: ?

IMaSC is a studio-recorded Malayalam TTS corpus: 49.63 hours and 34,473 text-audio pairs from 8 speakers (4 male, 4 female), at least 4.37 hours per speaker. Single-speaker VITS models trained on it reach an average naturalness MOS of 4.50, against 4.68 for ground truth.

Problem

Malayalam is agglutinative, so its vocabulary is very large and word lengths vary widely. Neural TTS therefore needs large, phonetically rich corpora for it. The public Malayalam TTS corpora that exist are small or have few speakers. IndicTTS (Baby et al., 2016) has 17.89 hours from 2 speakers. IndicSpeech has 29.1 hours from a single female speaker, and its authors attribute their weaker Malayalam MOS (compared with Hindi and Bengali) partly to insufficient vocabulary coverage. He et al. (2020) has 5.51 hours spread over 42 crowdsourced speakers, which is under 8 minutes per speaker. IndicVoices-R has more Malayalam audio, but it is derived from an ASR corpus, only about 5 hours of it is read speech, and its recording quality is lower than studio TTS data (§2.1, §5.3).

Method

The paper’s main contribution is the corpus. Text comes from Malayalam Wikipedia, which the authors chose because it is public domain. They started from 488,249 sentences, kept only sentences made entirely of Malayalam characters and punctuation, and had linguists from the University of Kerala vet the samples for naturalness, semantics and syntax. Sentences were selected so that every Malayalam phoneme appears at least 100 times. This produced 8,853 unique sentences. Speakers read overlapping subsets of these sentences, giving 34,473 recorded pairs in total (§3.1, §5.1).

Speakers were chosen subjectively for correct pronunciation, rhythm and consistency in the standard dialect, and to cover a range of prosody and articulation styles (§3.2, Table 1; ages 22 to 48). Recording took place in a soundproof studio in the University of Kerala Phonetics Lab with a RODE microphone, at 48 kHz, 24-bit, mono. The noise floor was below -60 dBFS and speech was kept at -20 dBFS or above. A monitor flagged misreadings during sessions and those sentences were re-recorded (§3.3). Audio was segmented automatically at silences and then corrected by hand. Language experts checked each audio-text pair to phoneme level: pairs with substantial mismatches were discarded, and pairs with slight mismatches had the text corrected to match the audio (§3.4).

To validate the corpus, the authors trained one VITS model per speaker using the Coqui TTS implementation. The input is raw Malayalam characters, with no G2P front end, since Malayalam orthography is close to phonemic. Each model was trained for 60k steps on a single Tesla P100 (§4, §5.2). VITS is an end-to-end conditional VAE with adversarial (GAN) waveform training, which is why the architecture is tagged VAE and GAN. Model size is not reported.

Key Results

Corpus statistics (Table 2). Per-speaker durations range from 4:21:56 (F3) to 9:34:21 (F4). F4’s larger share comes from extra experimental sessions. The text contains 225,781 words in total, 23,604 of them unique, and 1,911,388 phonemes. Most utterances last 2 to 8 seconds (Figure 3). Text length in characters correlates roughly linearly with duration, and word count correlates more loosely, which reflects agglutination (Figures 4 and 5).

Comparison with other Malayalam corpora (Table 4). IMaSC is the largest of the four studio or read-speech Malayalam corpora compared, on sentences (34,473 vs. 19,954 for IndicSpeech), words (225,781 vs. 109,245) and hours (49.63 vs. 29.1). He et al. (2020) has more speakers (42 vs. 8) but much less audio per speaker. The comparison covers size only. No TTS model was trained on the other corpora, so it says nothing about relative recording or transcription quality.

MOS (Table 3). Twenty native speakers, recruited by crowdsourcing, rated naturalness on a 5-point scale. Each rater heard 10 synthesized sentences and 2 ground-truth recordings per speaker (§5.2.1). Synthesized MOS ranges from 4.38 ± 0.11 (M2) to 4.60 ± 0.09 (M1), averaging 4.50 ± 0.04. Ground truth ranges from 4.55 to 4.90, averaging 4.68 ± 0.08. The per-speaker gap between synthesis and ground truth is between 0.04 (F4) and 0.32 (F3). The speaker with the most data has the smallest gap and the speaker with the least data has the largest. However, each ground-truth score rests on only 2 sentences (95% intervals up to ±0.26), so this pattern is suggestive at most. The paper reports no objective metrics (WER, CER, speaker similarity, or an automatic MOS predictor).

Novelty Assessment

The contribution is data, not method. The recording pipeline (phonetically balanced selection from Wikipedia, studio recording with live monitoring, expert phoneme-level verification) and the evaluation protocol (an off-the-shelf VITS model per speaker plus a MOS test) are standard for TTS corpus papers. VITS itself is unchanged. What is new is the scale and speaker count for Malayalam: about 1.7 times the hours of the previous largest public studio corpus, with 4 male and 4 female voices that each have enough data for a single-speaker model. The evaluation shows that the corpus is sufficient for good single-speaker synthesis. It does not show that IMaSC is better than existing corpora in quality, and it does not test the multi-speaker use the conclusion points to.

Field Significance

Moderate. IMaSC is the largest studio-quality, multi-speaker public TTS corpus for Malayalam reported in the paper’s comparison, with at least 4.37 hours per speaker. It can serve as training data for single-speaker and multi-speaker Malayalam synthesis and for Dravidian-language TTS evaluation. Its evidence is limited to showing that a standard end-to-end model trained on each speaker’s subset reaches MOS within about 0.2 of ground truth.

Claims

  • supports: In languages with near-phonemic orthography, end-to-end neural TTS trained directly on character input, without a separate grapheme-to-phoneme front end, can reach naturalness close to recorded speech given several hours of studio data per speaker.

    Evidence: Per-speaker VITS models trained on tokenized raw Malayalam characters for 60k steps average 4.50 ± 0.04 MOS, against 4.68 ± 0.08 for ground truth, across 8 speakers with 4.37 to 9.57 hours each. (§5.2, §5.2.1, Table 3)

  • supports: A studio-quality single-speaker TTS corpus needs several hours per speaker, so crowdsourced multi-speaker corpora with only minutes per speaker cannot on their own support single-speaker voice building.

    Evidence: The authors argue that He et al. (2020), with 5.51 hours over 42 speakers (under 8 minutes each), makes single-speaker TTS infeasible, whereas every IMaSC speaker has at least 4.37 hours (average 6.20) and supports a model rated 4.38 to 4.60 MOS. (§5.3, Table 2, Table 4, Table 3)

  • complicates: Corpus papers that validate data with a single high MOS score from a model trained only on the new corpus show that the data is sufficient, not that it is better than existing corpora.

    Evidence: Table 4 compares IMaSC with other Malayalam corpora on size and speaker count only. No model is trained on IndicSpeech or the other corpora under the same protocol, and no objective intelligibility metric is reported. The ground-truth MOS for each speaker rests on 2 sentences. (§5.2.1, §5.3, Table 3, Table 4)

Limitations and Open Questions

No comparative or objective validation

The corpus is validated only by MOS for VITS models trained on IMaSC itself. No model is trained on competing Malayalam corpora, and no WER, CER or other objective intelligibility measure is reported. The results therefore show that the corpus is usable, not that its quality beats the alternatives.

  • The paper presents the corpus as suited to multi-speaker TTS, but no multi-speaker model is trained or evaluated. Speaker separability and cross-speaker effects are untested.
  • The MOS test uses 20 crowdsourced raters and 10 synthesized sentences per speaker. Ground truth is scored on only 2 sentences per speaker, so its intervals are wide (±0.14 to ±0.26), which limits per-speaker comparisons.
  • The parsed text does not give a test-set split procedure, a release URL or a licence. Wikipedia sourcing is described as public domain, but the terms for the audio are not stated.
  • The sentence set is shared across speakers (8,853 unique sentences for 34,473 recordings), so the corpus has less lexical diversity than its total pair count suggests. Text coverage is limited to encyclopedic Wikipedia prose, with no conversational or expressive material.
  • Speakers were selected subjectively for “pleasing” articulation in the standard dialect, so dialectal variation in Malayalam is not represented.

Wiki Connections

  • GAN Vocoder — the validation models are VITS, whose end-to-end waveform generation is trained adversarially, and the paper applies it unchanged to a new low-resource language.
  • Subjective Evaluation — corpus quality is judged mainly by a 20-rater, 5-point naturalness MOS test that compares synthesized speech with ground truth for each speaker.
  • Tacotron 2 — cited as one of the two-stage architectures used to validate earlier TTS corpora; the authors chose VITS instead.
  • Tacotron — cited alongside Tacotron 2 as the architecture used to evaluate corpora such as CSS10 and KazakhTTS.
  • LibriTTS — cited as an earlier TTS corpus validated by training neural TTS models, the evaluation pattern IMaSC follows.
  • A Survey on Neural Speech Synthesis — cited to motivate the shift away from two-stage acoustic-model-plus-vocoder pipelines toward end-to-end synthesis.