arXiv · 2025 · Preprint
Irengbam et al. · → Paper · Demo: ? · Code: ?
This paper applies Tacotron 2 and HiFi-GAN to build the first TTS system for Manipuri written in the Meitei Mayek script, accompanied by a custom phoneme mapping from Meitei Mayek characters to ARPAbet.
Problem
Manipuri (Meitei Lon), one of India’s 22 scheduled languages, has virtually no presence in speech technology. Prior neural TTS work has focused on Indo-Aryan and Dravidian languages with Latin or Brahmi-derived scripts; the Meitei Mayek script — a revived indigenous writing system — has no published TTS corpus, no phoneme table for standard synthesis toolkits, and no end-to-end system. Additionally, Manipuri is tonal, with at least two contrastive tones, a phonological property that most existing TTS pipelines for Indian languages do not address.
Method
The system follows the standard Tacotron 2 + HiFi-GAN two-stage pipeline. Input Meitei Mayek text is first normalised (cleaning, abbreviation expansion, standardisation) and then converted to phonemes through a manually constructed lookup table mapping each of the 55 unique Meitei Mayek characters to its closest ARPAbet equivalent. The phoneme sequence is passed to a Tacotron 2 acoustic model — encoder (3 CNN layers + BiLSTM), location-sensitive attention, LSTM decoder, and a 5-layer 1D-CNN postnet — which predicts an 80-band mel-spectrogram. HiFi-GAN then converts the spectrogram to a waveform via adversarial training.
Training data consists of 818 utterances (~40 minutes) recorded by a single native speaker at 22.05 kHz. The 80:10:10 train/validation/test split is applied at the utterance level. The model was trained for 310 epochs with Adam, an initial learning rate of 0.001 with exponential decay, and batch size 32. Training loss stabilises below 0.08 by the final epochs. No cross-lingual transfer or pre-training on other languages is used; the Tacotron 2 is trained from scratch on the Meitei Mayek corpus.
Key Results
Evaluation is conducted through subjective MOS ratings collected from native speakers. The system achieves a naturalness MOS of 3.34 ± 0.70, pronunciation accuracy MOS of 3.51 ± 0.31, and an overall MOS of 3.43 ± 0.51. No automated metrics (WER, SPK-SIM) are reported, and no baseline system exists for direct comparison — this is the first reported TTS result for Meitei Mayek, so all numbers are self-assessed. Evaluators indicated output was intelligible in most cases; the main failures were on rare phoneme sequences and long, syntactically complex sentences, where rhythm degraded.
Note
No prior system exists for comparison, so reported MOS values reflect absolute quality on this script/language only. Generalisation beyond the single training speaker and the narrow test set is untested.
Novelty Assessment
The contribution is engineering integration: Tacotron 2 and HiFi-GAN are standard, well-established components, applied here to a new language and script. The genuine novelty is the ARPAbet phoneme mapping for Meitei Mayek — a necessary prerequisite for any future TTS work on this script — and the dataset itself, however small. The paper does not introduce new architectures, training objectives, or methods for handling tonal phonology, which the authors acknowledge as a limitation. The tonal contrast (falling vs. level tone) present in Meitei Mayek is not explicitly modelled; tone markers are presumably consumed as character symbols in the phoneme table without dedicated acoustic conditioning.
Field Significance
Low — This paper’s contribution to speech synthesis methodology is minimal: it applies a circa-2018 pipeline to a new script. Its significance is specific to the Manipuri NLP and language preservation community, where it establishes the first existence proof for Meitei Mayek TTS, provides the only known Meitei Mayek phoneme-to-ARPAbet mapping, and releases the only known (if small) aligned speech corpus for the script. For the broader TTS field, it offers confirmatory evidence that standard neural TTS pipelines can produce intelligible output even from very small (40-minute) single-speaker corpora of previously unsupported scripts.
Claims
- Standard neural TTS pipelines (acoustic model + neural vocoder) can produce intelligible speech for previously unsupported tonal scripts with as little as 40 minutes of single-speaker training data. (§5.3.5, §5.4)
- The absence of explicit tone modelling in a Tacotron 2 pipeline leads to degraded prosody on complex utterances in tonal languages, even when the training corpus is small. (§2.1, §5.4)
- Developing a phoneme mapping between a novel script and an established phoneme set (such as ARPAbet) is a necessary and non-trivial prerequisite before standard TTS toolkits can be applied to indigenous writing systems. (§3.1.4, §6.2)
Limitations and Open Questions
Warning
The system is trained on a single speaker with only 40 minutes of audio, and no comparison to any baseline is possible. MOS scores are self-contained and cannot be situated relative to other low-resource systems.
Tonal phonology is not explicitly modelled. Manipuri distinguishes at least two contrastive tones, and the falling tone is marked orthographically in Meitei Mayek, but the paper does not describe any mechanism to condition acoustic output on tone — the tone diacritic is treated as another character in the phoneme table. This likely explains the reported flatness on longer utterances.
The dataset is very small (818 utterances, one speaker) and remains unreleased at time of publication, limiting reproducibility. The practical path forward requires more data, speaker diversity, and either explicit prosody conditioning or fine-tuning from a related-language pre-trained model.
Wiki Connections
This paper is most directly relevant to subjective-evaluation and evaluation-metrics. The use of HiFi-GAN as vocoder connects to gan-vocoder.