Kashmiri Speech Technology: Datasets for TTS and Speech Recognition

Kashmiri speech technology is at an early stage, but as of October 2026 there are open datasets and models for both directions: Bolbosh, an open-source text-to-speech system with a roughly 80-hour training corpus, and AI4Bharat's IndicConformer model for speech recognition, built alongside the IndicVoices corpus. Mainstream tools lag behind: OpenAI's Whisper does not list Kashmiri, and Meta's MMS covers it only for language identification.

This post lists the speech resources we could verify, what each contains as its creators describe it, and the licence attached. The wider list of Kashmiri datasets and tools is in our Resource Hub.

Where Kashmiri speech technology stands

Until recently there was little to work with beyond small academic collections. Three things changed that. AI4Bharat's multilingual speech projects (IndicVoices, IndicVoices-R and Rasa) each include a Kashmiri portion. In March 2026 a team from GAASH Lab at NIT Srinagar, the University of Kashmir and KAUST released Bolbosh, which describes itself as the first dedicated, open-source neural text-to-speech system for Kashmiri. And India's AIKosh platform now hosts Kashmiri audio, including a single-speaker TTS set and a month of radio news.

The numbers are still modest. The Bolbosh authors report that the best Kashmiri speech recogniser they tested had a word error rate of 41.20% even after removing vowel marks, and conclude that “Kashmiri ASR remains underdeveloped”. Anyone building a Kashmiri voice product should treat these resources as starting points, not finished components.

Resources at a glance

Sizes are given as each creator states them. Several entries are portions of multilingual collections, so check the dataset card for the Kashmiri subset before planning around a figure.

Text-to-speech datasets and models

Bolbosh

Bolbosh: Script-Aware Flow Matching for Kashmiri Text-to-Speech (Tajamul Ashraf, Burhaan Rasheed Zargar, Saeed Abdul Muizz, Ifrah Mushtaq, Nazima Mehdi, Iqra Altaf Gillani, Aadil Amin Kak and Janibul Bashir; arXiv, March 2026) adapts the Matcha-TTS architecture, starting from an English multi-speaker checkpoint. The authors expand the model's input vocabulary to 272 graphemes so that Kashmiri letters and diacritics are represented explicitly.

Their corpus totals 79.89 hours: 36.28 hours of studio recordings from AI4Bharat's Rasa and 43.61 hours from IndicVoices-R, the latter used only for training and cleaned with a denoising and loudness-normalisation pipeline. Validation and test sets come from Rasa alone. In a listening test with 32 native Kashmiri speakers, the authors report a mean opinion score of 3.63 for Bolbosh, against 1.86 for the multilingual IndicParler baseline and 4.61 for real recordings. The curated corpus is on Hugging Face (39,996 utterances split 33,182 / 4,542 / 2,272), the code is on GitHub under MIT, and the project page links a demo.

Rasa and IndicVoices-R

Both are AI4Bharat datasets covering the 22 scheduled languages of India, with a Kashmiri configuration in each. Rasa is a studio-recorded expressive TTS corpus; IndicVoices-R is built from mostly extempore IndicVoices speech, enhanced for TTS use. Both are CC BY 4.0 and require you to share contact details on Hugging Face before downloading.

KTTS Single Speaker Dataset

A smaller, single-voice set by Kh Mohmad Shafi, Asif Ali Bhat, Kamran Imtiyaz and Javaid Iqbal: 2,984 WAV recordings at 48 kHz from one male speaker, with the matching text in a textcorpus.csv file. It appears as version 2 on Mendeley Data (published 29 August 2024) under CC BY-NC 3.0, and on AIKosh, where the listing gives CC BY 3.0. The two licences differ; until that is resolved, treat it as non-commercial. A single clean voice is a reasonable fit for training a one-speaker synthesiser or for fine-tuning experiments.

Speech recognition data and models

IndicVoices and IndicConformer

IndicVoices (Javed et al., Findings of ACL 2024) is AI4Bharat's large collection of read, extempore and conversational speech. Its dataset card reports 23.7K hours, 11.2K of them transcribed, across 22 languages, and Kashmiri is one of its configurations. It is CC BY 4.0 and gated behind a contact-details form.

AI4Bharat also publishes a Kashmiri-specific recogniser, IndicConformer (ks): a 120M-parameter Conformer-Large with hybrid CTC and RNN-T decoders, MIT-licensed, which takes 16 kHz mono audio and runs on AI4Bharat's fork of NVIDIA NeMo. The model card does not report a word error rate; the Bolbosh paper's figures (below) are the only independent ones we found.

Smaller and older collections

• SLR122, the Kashmiri Data Corpus: transcribed recordings of native speakers, collected by a group of students from Kashmir for an ASR project. GPL-3.0-or-later; processing scripts are in the kscp repository.
• Kashmiri Audio Segmented Corpus: 1,955 short segments sliced automatically from SLR122, with transcripts and durations. It inherits GPL-3.0.
• Frequently Spoken Kashmiri Words: 12 words recorded from native speakers across three regions of Kashmir, for keyword and voice-command research.
• Akashvani Kashmiri news, November 2025: radio bulletins and advisories in MP3 with no transcripts. It could support transcription work, but AIKosh lists its licence as “Not applicable”, so confirm reuse terms before training on it.

Whisper and MMS

OpenAI's Whisper does not include Kashmiri: there is no ks entry in the language list in its tokenizer, although Urdu and Sindhi are there. Forcing Whisper to transcribe Kashmiri as Urdu will produce Urdu-like text, not Kashmiri orthography.

Meta's Massively Multilingual Speech project covers more than 1,100 languages for ASR and TTS, but its language coverage table marks Kashmiri (kas) for language identification only, with no ASR or TTS model. MMS weights are CC BY-NC 4.0. The Bolbosh authors also tested Meta's newer OmniASR models on Kashmiri and report word error rates between 87.67% and 94.34%.

Script and vowel-mark issues in speech data

Kashmiri in Perso-Arabic script marks several vowels with diacritics, for example in کٲشُر. Many writers leave some of these marks out, so the same word can appear in several spellings across a transcript corpus. For background on the scripts and their vowel signs, see our post on the history of Kashmiri scripts.

This matters in both directions, as the Bolbosh paper shows:

• Speech recognition. IndicConformer's RNN-T decoder scored a 66.59% word error rate when diacritics were kept in the comparison and 41.20% when they were stripped. Much of the measured error is about vowel marks, not wrong words, so always state which convention an ASR score uses.
• Text-to-speech. The authors trace the multilingual baseline's poor score to frequent vowel mispronunciations and weak handling of Perso-Arabic diacritics. With diacritics in the input text, Bolbosh's relative word error rate was 4.14%; without them it rose to 13.23%.

The practical advice is to normalise Unicode variants consistently, keep diacritics where the source has them, and record which normalisation you applied. The same caution applies to the Devanagari Kashmiri used in some sources (कॉशुर): the resources above that we checked use Perso-Arabic transcripts, so mixing scripts needs a deliberate conversion step.

Licences

• CC BY 4.0 (Bolbosh corpus, Rasa, IndicVoices, IndicVoices-R): commercial use allowed with attribution. AI4Bharat's sets are gated, so each user must accept the terms.
• CC BY-NC 3.0 (KTTS single-speaker set on Mendeley; its AIKosh listing says CC BY 3.0): attribution required, and non-commercial use only if you follow the stricter of the two.
• MIT (IndicConformer model, Bolbosh code, the spoken-words set): permissive.
• GPL-3.0 (SLR122 and the segmented corpus derived from it): a copyleft licence written for software. Get advice before you ship models or derived data built from it in a product.
• CC BY-NC 4.0 (Meta MMS): non-commercial only.
• Unclear (Akashvani broadcast set): ask the publisher before reuse.

Bolbosh combines Rasa and IndicVoices-R audio, so credit AI4Bharat's sources as well as the Bolbosh authors when you use it.

How to contribute recordings

Kashmir AI Research does not collect speech recordings; our own work is on text (see the Kashmiri NLP resource guide for text resources). If you want to add Kashmiri voices to open data, these are the public routes we could confirm:

• Mozilla Common Voice. Common Voice's language list includes Kashmiri (ks), but as of October 2026 its public API marks it as not yet open for contributions. Speakers who want to help can follow it from the Common Voice languages page and check there when recording opens.
• Publish your own data. If you record with consent, publishing on Hugging Face or Zenodo with a clear licence, a speaker-consent statement and a note on the script and diacritic conventions you followed makes the data far more useful than a bare archive.

If you know of a Kashmiri speech resource missing here, the Resource Hub is where we track additions.

Sources

• Ashraf et al. (2026), Bolbosh: Script-Aware Flow Matching for Kashmiri Text-to-Speech, arXiv:2603.07513; dataset; code
• AI4Bharat, IndicVoices, IndicVoices-R and Rasa on Hugging Face; IndicVoices project page
• AI4Bharat, IndicConformer Kashmiri model card
• KTTS Single Speaker Dataset, Mendeley Data, v2; AIKosh listing
• Kashmiri News Broadcast, Nov 2025 (Akashvani), AIKosh
• OpenSLR SLR122: Kashmiri Data Corpus
• OpenAI, Whisper language list; Meta, MMS language coverage
• Mozilla Common Voice languages