Whisper-Hindi2Hinglish is an open-source speech recognition model we fine-tuned from OpenAI's Whisper large-v3 on roughly 550 hours of noisy, Indian-accented Hindi audio. On three public benchmarks it roughly halves word error rate against the base model, and it has passed 190,000 downloads in the last month. It is released under Apache 2.0, and it is not our production speech stack.
What the model is
Whisper-Hindi2Hinglish is a fine-tune of OpenAI's Whisper large-v3, trained to transcribe Indian speech into spoken Hinglish rather than forcing it into either Devanagari Hindi or English. It is published on Hugging Face under an Apache 2.0 licence, so anyone can use it commercially without asking us.
- Base model: openai/whisper-large-v3, 2 billion parameters.
- Training data: roughly 550 hours of noisy, Indian-accented Hindi audio, curated in-house because no ASR-ready Hinglish dataset existed to use.
- Labelling: machine-transcribed first, then corrected by people.
- Deliberately noisy: the audio was chosen for background noise, because that is the condition the model has to survive in India.
- Licence: Apache 2.0. Downloads in the last month: 190,811 (Hugging Face, 20 August 2026).
Why Hinglish breaks generic speech recognition
Most speech models are trained on clean, single-language audio. Indian phone calls are neither. A caller starts a sentence in Hindi, finishes it in English, and does it over a patchy mobile line with a horn going off nearby. A model that has only seen tidy monolingual speech does not degrade gracefully on that. It invents.
The failure is not subtle. Put a real call through base Whisper and you get confident nonsense: Thank you, Sanchez for Kitne saal ki?, or Rangers, I can tell you for Lander cycle chaahie. It is not mishearing a word here and there. It is producing fluent English that has nothing to do with what was said, which is far more dangerous in a regulated conversation than an obvious error.
A speech model that fails loudly is a nuisance. One that fails fluently, in confident English, is a compliance problem.
The numbers, and how to read them
Word error rate on three public benchmarks, base model against the fine-tune. Lower is better. These figures are self-reported on our model card and Hugging Face labels them as such, so treat them as our measurement rather than an independent audit. The benchmark data and the code are public, so anyone can rerun them.
The honest reading: the fine-tune roughly halves error rate on Common Voice and cuts it substantially on FLEURS, with a smaller gain on Indic-Voices, which is the hardest of the three. A 60% word error rate on Indic-Voices is still a high number. We publish it because leaving it out would make the other two look better than they are.
| Whisper large-v3 | Whisper-Hindi2Hinglish-Prime | |
|---|---|---|
| Common Voice | 61.94 | 32.43 |
| FLEURS | 50.84 | 28.68 |
| Indic-Voices | 82.56 | 60.82 |
What this model is not
This is the part most vendors would leave out, so let me be direct about it.
This model is not our production speech stack, and speech recognition is not what we sell. The speech components in our platform are third-party and deliberately vendor-agnostic, so a client who mandates a specific voice or vendor can have one. We treat the speech layer as something anyone can buy, because increasingly it is.
What we actually defend sits above the speech: eight years of real Indic contact-centre interaction data, the orchestration and compliance layer, and the outcome the conversation is supposed to produce. A vendor leading with proprietary speech models as its headline differentiator is answering a question buyers stopped asking.
Because Indic ASR was underserved and we had the data to improve it. Open-sourcing a fine-tune costs us nothing competitively, since the model is not the moat, and it gives the people building Indian voice products a better starting point than base Whisper. That is the whole reasoning.
Open-sourced
- A Whisper fine-tune for Hinglish
- Weights, code and benchmarks public
- Apache 2.0, no strings
- Not our production speech stack
What we defend
- Eight years of Indic interaction data
- Orchestration and compliance layer
- The outcome the call produces
- Speech kept vendor-agnostic
How to use it
It loads through the standard Transformers pipeline, so if you have used Whisper you already know how to run it.
- Model card and weights: Whisper-Hindi2Hinglish-Prime on Hugging Face.
- A newer variant, Apex, performs better again.
- Compare it against other models on our open Speech-to-Text Arena.
- Code and conversion scripts: GitHub.
- Questions about the model itself go to the AI team, not to sales.
If what you actually need is the layer above the speech, the part that runs a collections or renewals conversation to an outcome, that is a different conversation and it starts on a live call.
Frequently asked questions
Is Whisper-Hindi2Hinglish free to use commercially?
Yes. It is released under an Apache 2.0 licence, so you can use it in commercial products without a separate agreement with us. The weights, the model card and the conversion code are all public.
What is it actually trained on?
Roughly 550 hours of noisy, Indian-accented Hindi audio, curated in-house because no ASR-ready Hinglish dataset existed. The audio was transcribed by a model first and then corrected by people. Noise was selected for rather than filtered out, because Indian call environments are noisy and a model trained on clean speech falls over on them.
How much better is it than base Whisper?
On our own benchmarking it averages roughly 39% lower word error rate across three public datasets, roughly halving error rate on Common Voice. Those figures are self-reported and labelled as such on Hugging Face. The datasets are public, so you can verify them yourself, and the Speech-to-Text Arena lets you compare it against other models on real audio.
Does Oriserve build its own speech recognition for production?
No. This model is an open-source contribution, not our production stack. The speech components in the platform are third-party and deliberately vendor-agnostic, so they can be swapped when a client mandates a specific vendor. What we defend is the outcome layer above the speech and the eight years of Indic interaction data behind it.
Which languages does it handle?
Hindi and English, including the code-switched Hinglish that sits between them, which is the case generic models handle worst. It is a single-purpose model for that mix rather than a general multilingual model.