Non-Profit Adalat AI Releases Open Indic ASR Models for Document-Ready Transcription

Legal tech non-profit Adalat AI has released Koyal, a family of five open speech recognition models for Hindi, Kannada, Malayalam, and Telugu. They are designed to generate document-ready transcripts with punctuation and formatted numerals rather than producing a raw stream of recognised words.

The release includes four 120-million-parameter monolingual models for Hindi, Kannada, Malayalam, and Telugu, and a 600-million-parameter multilingual model covering all four languages. The 120M models are designed for offline transcription, while the multilingual model supports both offline and streaming transcription from a single checkpoint.

The 600M multilingual model is also positioned as one of the earliest open cache-aware streaming ASR models supporting Malayalam, Kannada, and Telugu in native scripts from a single checkpoint.

Most open automatic speech recognition (ASR) systems are primarily designed to recognise words. Developers often have to add separate systems for punctuation restoration, number formatting, and language-specific post-processing before the transcript can be used as a document.

Koyal is designed to address that gap at the model level.

The models are trained to produce what Adalat AI calls “rich orthography”, retaining punctuation and written number forms in the transcript. This is particularly relevant for applications where the transcript itself is the final output, including dictation, legal documents, forms, reports, and notes.

“Recognition alone is not the product,” Adalat AI stated in a LinkedIn post, describing the problem it has encountered while building speech-to-document workflows.

The four monolingual models are approximately 0.5 GB each, making them suitable for deployments with more limited hardware resources.

The monolingual models are based on Conformer hybrid TDT-CTC systems from AI4Bharat's MahaDhwani encoder, which was pre-trained on 279,000 hours of Indic speech. They were subsequently fine-tuned on 1,350–2,400 hours of curated dictation data per language.

The multilingual model uses a cache-aware FastConformer encoder adapted from NVIDIA's Nemotron 3.5 ASR, along with a tokenizer and decoder designed for Hindi, Kannada, Malayalam, and Telugu.

On Adalat AI's FLEURS-RO rich-orthography benchmark, the 120M Koyal models recorded WER_SCRIBE scores of 9.03% for Hindi, 11.30% for Kannada, 14.74% for Malayalam, and 17.81% for Telugu.

The company said the models reduce error rates by roughly half compared with its earlier Whisper-small releases while using around half the parameters for the three languages covered by the comparison.

The benchmark differs from conventional word error rate because it evaluates not only word recognition but also punctuation and numeral rendering—the elements Adalat AI considers necessary for document-ready transcription.

The 600M multilingual model targets applications that require multiple Indic languages or real-time transcription.

The model supports different inference settings based on future audio available to the system. Adalat AI's tests show that at zero lookahead, the model remains within 2.22 percentage points of full-context offline decoding across all four languages, indicating a relatively gradual accuracy trade-off as latency is reduced.

The models are being released openly through Hugging Face, with Adalat AI providing model cards containing information on training data, evaluation, and deployment.