How speech to text can run without the cloud

The pipeline, end to end

When you dictate, Veena records your voice as 16 kHz mono audio. A preprocessing step converts that raw audio into the feature representation the model expects. The features go into a neural network that outputs, for every slice of time, a probability over possible sound units. A decoder turns those probabilities into words, and the words appear in the transcript box. All of these steps run as local code on your phone or computer.

The model: IndicConformer

Veena's engine is built on IndicConformer, a speech recognition model published by AI4Bharat, a research lab at IIT Madras that works on Indian-language datasets and models. IndicConformer was trained on speech from all 22 scheduled languages of India, which is why a single model can handle Telugu and Bodo without being retrained.

The architecture is a Conformer, a network designed for audio: it combines a convolution module, which is good at local sound patterns, with self-attention, which tracks long-range context across the sentence. It is trained with a CTC objective, which lets the model align sounds to letters without anyone labelling exactly where each word starts and ends.

One multilingual model, 22 language masks

Here is the part that surprises people: Veena does not ship 22 separate models. The multilingual model has one shared output head, and each language owns a slice of it. When you pick Kannada, the decoder only considers the classes in the Kannada slice; pick Bengali and it only considers Bengali classes. That masked decoding is why the model stays about the same size whether you use one language or all of them, and why a sentence in one language does not drift into another.

Telugu is the exception worth knowing about: it gets its own dedicated model, tuned specifically for the language.

Making it fit on a phone: quantization

A research model like this is large, and phones are not servers. Veena runs the model exported to ONNX, a standard format for shipping neural networks, and quantized to INT8: each weight is stored as an 8-bit integer instead of a 32-bit float. The model gets several times smaller and runs faster, at a small cost in precision that careful calibration mostly recovers. This makes local transcription practical. The time it takes depends on your device and the length of your recording.

Why offline changes the experience

• Latency. There is no upload and no server queue, so the transcript appears as soon as the model finishes, which is faster than a network round trip on a bad connection.
• Availability. Airplane mode, a basement, a village with two bars of signal: none of it matters.
• Privacy. The audio is never sent anywhere, so there is no server log to leak and no retention policy to trust.

Where it still needs the network

Once: the first launch, to download the model itself, about 125 MB for Telugu or about 480 MB for the full multilingual model. After that download completes, you can turn off wifi and data indefinitely. If you want the product view rather than the technical one, see the feature list, or read how speech to text works in general on the blog.