- The paper introduces Bolbosh, an open-source Kashmiri TTS system that adapts Matcha-TTS with optimal-transport flow matching, expanded grapheme coverage, and internal alignment.
- Bolbosh achieves a 3.634 MOS and 3.73 MCD, substantially outperforming zero-shot IndicParler, which records a 1.864 MOS and 4.73 MCD on the same task.
- The study shows that preserving Kashmiri diacritics is crucial for intelligibility, with relative WER increasing from 4.14% to 13.23% when they are removed, while also highlighting gaps in dialect and spontaneous-speech evaluation.
Motivation and problem setting
Kashmiri, a Dardic Indo-Aryan language spoken by roughly 7 million people, has no publicly available neural text-to-speech (TTS) system despite its official status. The paper identifies three compounding obstacles: scarce paired text–speech corpora, multi-script orthography (Perso-Arabic, Devanagari, Roman) with diacritics that carry essential vowel distinctions, and substantial dialectal variation. The authors first quantify the inadequacy of zero-shot multilingual transfer: IndicParler, a recent Indic multilingual TTS framework, achieves a Mean Opinion Score (MOS) of only 1.86 on Kashmiri, with frequent vowel mispronunciations and prosodic distortion. The paper attributes this failure to inadequate handling of Perso-Arabic diacritics and phonotactic mismatch with high-resource training languages — a claim with implications beyond Kashmiri, since it suggests multilingual TTS coverage is not merely a matter of language count but of script-aware supervision.
The Bolbosh system
Bolbosh is a supervised cross-lingual adaptation of Matcha-TTS, an Optimal Transport Conditional Flow Matching (OT-CFM) architecture. The OT-CFM decoder learns a continuous velocity field transporting a Gaussian prior to the mel-spectrogram distribution, avoiding the iterative sampling cost of diffusion models, the hyperparameter sensitivity of GAN-based systems such as VITS, and the external-aligner dependency of FastSpeech 2 — the latter being particularly relevant given the absence of Kashmiri G2P resources. Monotonic Alignment Search (MAS) is performed internally, and the model operates directly on graphemes.
The adaptation strategy has three components:
- Cross-lingual initialization: fine-tuning from a pretrained multi-speaker English Matcha-TTS checkpoint, which provides acoustic priors and accelerates MAS convergence.
- Vocabulary expansion: the grapheme vocabulary is extended to 272 symbols to explicitly encode Kashmiri characters and diacritics, with language-specific text cleaners disabled so the encoder learns end-to-end grapheme-to-acoustic mappings.
- Multi-speaker regularization: training mixes studio-quality RASA recordings (36.28 h) with enhanced spontaneous IndicVoices-R data (43.61 h), using learned speaker embeddings per utterance; inference conditions exclusively on RASA speaker embeddings.
A three-stage acoustic enhancement pipeline bridges the domain gap between spontaneous and studio data: dereverberation/denoising via Resemble-Enhance (UNet denoiser with latent CFM refinement), silence trimming below 40 dB peak amplitude, and loudness normalization to −23.0 LUFS with resampling to 22.05 kHz. Text normalization canonicalizes Unicode variants, expands numbers, and — critically — preserves pronunciation-critical diacritics rather than stripping them.
Evaluation methodology
The evaluation is notable for its honesty about Kashmiri ASR limitations. The authors benchmark proxy ASR systems before using them for intelligibility measurement: IndicConformer RNN-T achieves 41.20% WER without diacritics but degrades to 66.59% with them, and Meta's OmniASR models perform far worse (up to 94.34% WER). Because ground-truth ASR error is so high, the paper introduces a Relative WER (rWER) normalized against ASR performance on ground-truth recordings, and treats WER as supplementary. Objective fidelity uses MCD with DTW alignment; subjective quality comes from a listening study with 32 native speakers. Validation and test splits are drawn exclusively from the studio RASA corpus, keeping evaluation controlled — though this also means evaluation reflects studio conditions rather than spontaneous speech.
Results
The headline numbers are substantial. Bolbosh achieves an MCD of 3.73 versus 4.73 for IndicParler, and a MOS of 3.634 (±0.061) versus 1.864 (±0.065) for the baseline, against a ground-truth MOS of 4.614 (±0.059). The near two-MOS-point margin over the zero-shot baseline is the paper's strongest quantitative claim.
| System |
MOS |
MCD |
rWER (%) |
| Human (ground truth) |
4.614 ± 0.059 |
— |
— |
| Bolbosh |
3.634 ± 0.061 |
3.73 |
4.14 |
| IndicParler |
1.864 ± 0.065 |
4.73 |
46.75 |
The ablation on diacritics is the most instructive result: Bolbosh's rWER rises from 4.14% to 13.23% when diacritics are removed, while IndicParler collapses to 100.32% rWER (WER 0.8253) under the same condition. This directly supports the central claim that explicit diacritic modeling is decisive for intelligibility in this orthography, and that the baseline's grapheme-to-acoustic mapping is unstable without diacritic supervision. Spectrograms corroborate the metrics: Bolbosh preserves harmonic structure and formant trajectories, whereas IndicParler shows over-smoothing and temporal instability.
Limitations and open questions
Several constraints qualify these results. First, the 1.36-MOS gap to ground-truth recordings remains nontrivial, and the paper does not decompose it. Second, evaluation is confined to studio-quality RASA test data; performance on spontaneous, noisy speech — the condition most speakers actually encounter — is unmeasured, and the enhancement pipeline's contribution is not isolated through ablation. Third, the proxy ASR's 41.20% baseline WER means rWER values, while useful comparatively, rest on a noisy measurement instrument; the paper concedes this by demoting WER to supplementary status. Fourth, the comparison baseline is a single zero-shot multilingual system; no fine-tuned non-flow-matching alternative (e.g., adapted VITS or FastSpeech 2) is evaluated, so the specific advantage of OT-CFM over other architectures under identical data conditions is asserted rather than demonstrated. Finally, dialectal variation within Kashmiri is acknowledged as a challenge but not addressed; the model's behavior across dialects is an open question.
Conclusion
Bolbosh establishes the first open-source, script-aware neural TTS system for Kashmiri, combining supervised OT-CFM fine-tuning from an English multi-speaker checkpoint, diacritic-preserving grapheme vocabulary expansion, and corpus enhancement to unify heterogeneous acoustic domains. Its results — MOS 3.63 and MCD 3.73 against a 1.86-MOS zero-shot baseline — demonstrate concretely that zero-shot multilingual transfer is insufficient for diacritic-sensitive scripts and that modest supervised data with script-aware encoding yields near-natural synthesis. The paper leaves open the questions of spontaneous-speech robustness, dialect coverage, and controlled architectural comparisons, but it provides both a benchmark and a reproducible recipe for low-resource TTS in similarly under-served languages.