Building Kashmiri speech recognition on one server, without a GPU

By Sheikh Muhibul Hassan · 2026-09-10 · Srinagar, Jammu and Kashmir

Kashmiri is spoken by around seven million people and has almost no speech technology. There is no dictation for it in any keyboard, no voice assistant that understands it, and until recently no open model anyone could run themselves. This is what it took to get speech-to-text working for it on one ordinary server.

Does Kashmiri speech recognition actually work?

Yes, for transcription. Kashmiri and Urdu speech-to-text now runs self-hosted on a single six-core virtual server with no GPU, transcribing roughly six times faster than real time. It is built on AI4Bharat’s IndicConformer models, which are MIT licensed and officially support Kashmiri. Text-to-speech is a different story, covered below.

Which models exist for Kashmiri?

Two, both from AI4Bharat, both MIT licensed:

That is the whole list. For most of the world’s languages this section would be empty, and until AI4Bharat published these it was empty for Kashmiri too.

What the error rate actually says

Measured against eight sentences from the OpenSLR Kashmiri corpus in September 2026, word error rate came out at 73%. That number is close to meaningless, and understanding why matters more than the figure.

The corpus was transcribed by students writing Kashmiri the way they write it in practice. The model outputs standard orthography, with the diacritics that formal written Kashmiri uses and everyday writing usually drops. Word error rate compares strings, so two spellings of the same spoken word count as a miss. Most of that 73% is orthographic disagreement, not the model mishearing anything.

This is a general trap when evaluating speech recognition for a language whose written form is not settled. Before trusting a word error rate for any low-resource language, read the reference transcripts and check whether they and the model even agree on how to spell the words.

Why the voice agent replies in Urdu

Because Kashmiri text-to-speech does not exist. Not as an open model, not as a paid API, not from any major cloud vendor. The same gap turned up independently while researching voice systems for hospitals in Kashmir.

So a Kashmiri voice agent has to be honest about what it is: it listens in Kashmiri and answers in an Urdu voice. The two languages share a Perso-Arabic script and enough vocabulary that Kashmiri speakers follow Urdu comfortably. It is a real compromise rather than a solved problem, and worth stating plainly to anyone being sold one.

What broke along the way

AI4Bharat’s toolchain is a fork of NVIDIA NeMo, and the fork expects older versions of three packages than a current environment installs. The container would not start until numpy was pinned to 1.26, with lightning and pyarrow pinned alongside it. Three dependency conflicts, each surfacing only once the previous one was fixed.

The other surprise was the language model. A general-purpose model handling the conversation writes Kashmiri that a Kashmiri speaker cannot read, and misreads Kashmiri numerals. Speech recognition and language generation are not equally solved for a low-resource language, and a design that assumes they are will produce transcripts that are fine and replies that are nonsense.

What this is useful for

A front desk that takes a Kashmiri voice note and returns text. A hospital reception line for patients not comfortable in English or Hindi. A WhatsApp voice-note desk for a business whose customers speak Kashmiri and type nothing. None of these needed a GPU, and none of them needed a vendor.

Questions

Is there an open Kashmiri speech-to-text model?
Yes. AI4Bharat’s IndicConformer, MIT licensed, in a 600M multilingual version on Hugging Face and a 120M Kashmiri-only version on AIKosh. Both can be self-hosted.
Is there a Kashmiri text-to-speech model?
No. As of September 2026 no open model, paid API or cloud vendor offers Kashmiri speech synthesis. Systems that need to speak Kashmiri use an Urdu voice instead.
Do you need a GPU to run Kashmiri speech recognition?
No. It runs on a six-core virtual server with no GPU at roughly six times real time, fast enough for voice notes and front-desk transcription. Live conversational latency at scale would want a GPU or a hosted speech API.
Why is the reported word error rate so high?
Because the reference transcripts and the model disagree about spelling, not about what was said. The corpus uses everyday student spelling; the model outputs standard orthography with diacritics. Most of the measured error is orthographic.

Sheikh Muhibul Hassan is an AI systems engineer in Srinagar, Jammu and Kashmir, and the founder of Lumivon AI Studio. He builds Android applications, WhatsApp and n8n automations, custom software and multilingual voice agents. Also known as Sheikh Hassan, Sheikh Muhib and Muhibul Hassan.

More about him · LinkedIn · X · GitHub

Have something like this to build? Start a project →

← All writing

← Back to Lumivon