Akshara: State-of-the-Art Speech Recognition for Indic Languages

General-purpose ASR models have reached impressive milestones on clean, long-form dictation. But when deployed in live, cascaded voice AI systems — where users speak over compressed, lossy cellular networks from noisy environments — these generic models fail. They suffer from semantic hallucination, high latency, and catastrophic degradation on the short, contextless utterances that define real B2C calls.

Akshara fundamentally rethinks the acoustic and linguistic modeling pipeline for Indian languages. It is not a fine-tuned Whisper derivative or a multilingual wrapper. It is a ground-up architecture engineered for the hostile realities of enterprise telephony — and it delivers state-of-the-art accuracy across nine Indic languages, as the benchmarks below show.

Automatic Language Detection

Akshara identifies the spoken language on its own — no configuration required. Pass no hint and it auto-detects the language directly from the audio. Pass a single code like hi and it transcribes purely in that language. Pass several like hi en and it transcribes the code-mixed speech natively, keeping every word in its own script.

Keyword Boosting for Domain Vocabulary

Every industry speaks its own language. Drug names, product SKUs, legal clauses, and internal jargon rarely appear in general training data, so even strong models stumble on them. With Akshara you can pass a list of keywords at inference time — the model biases decoding toward those terms and locks onto them, adapting to a new domain in seconds with no retraining or fine-tuning.

The same mechanism generalizes to any vertical — banking, insurance, logistics, or e-commerce. By simply supplying the terms that matter, teams tailor Akshara to their domain on the fly, making it a practical drop-in engine for specialized, real-world deployments.

Akshara Multilingual Benchmark Suite

Comprehensive WER, CER, intent, and entity evaluation against leading ASR systems across five public and proprietary datasets. Bold values indicate the best performance using the current VSI benchmark inputs.

Language-Specific Breakdown

Overall Word Error Rate per language, macro-averaged across available rows from IndicVoices, Kathbath, NoBroker, and VAANI. English uses NoBroker and VAANI; Svarah is excluded from this cross-dataset view. Lower is better.

End-to-End Transcription Latency

In cascaded voice AI, every millisecond counts. Latency is measured as the time from the end of a user utterance to the first byte of transcription output returned over the wire — under real telephony network conditions. Lower is better.