Anuvani-LID

Anuvani-LID

Identifies which of India's 22 scheduled languages, or Indian English, is spoken in an audio clip, with calibrated scores. For developers building multilingual speech pipelines.

Made by CIIL on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Anuvani-LID listens to a recording and says which language is spoken in it. Its label set is the 22 officially scheduled Indian languages, with Indian English as a 23rd. The model is a Conformer with about 116M parameters. Training used 54,643 hours of Indian speech, and the scores are tuned so you can read them as confidence.

Who it suits: It suits developers who build speech pipelines for Indian languages and need a first step that picks the language. That step decides which recognizer to run, which way to translate and which language to reply in.

Title: Who uses it
Industry | Who uses it | How they use it | Example
Speech technology | Pipeline developers | Route audio to the right recognizer | Pick the speech recognizer for a call recording
Customer service | Call-center engineers | Detect a caller's language early | Reply in the language the caller speaks
Mobile apps | App developers | Run language detection on the device | Identify a spoken language with no server

People who work only with languages outside India can skip it.

Where to use it:

• Python with transformers and torch
• NVIDIA GPUs behind Triton Inference Server
• CPUs through ONNX Runtime
• Phones and ARM Linux through sherpa-onnx, for Android, iOS and ARM Linux
• Google Colab, through the notebook

Title: Top jobs
Use case | Who | What they get back
Language detection | Pipeline developers | The spoken language, with a calibrated score
Long audio | Audio engineers | A timeline of languages and each language's share of the speech
Rejecting non-speech | Product teams | Music and non-Indian languages flagged as out of scope

How to set it up:

• Python on your computer
• The huggingface_hub package
• torch and transformers, the only runtime dependencies for the PyTorch model

1. Install the Hugging Face hub package.
2. Download the model files to your computer.
3. Load the model through the transformers pipeline or AutoModel, as in the card's notebook.

pip install -U huggingface_hub
hf download CIIL/Anuvani-LID

How to use it: Load the model with the transformers pipeline or with AutoModel, then pass in an audio clip. For recordings of any length, a helper function finds the speech and reports the languages along a timeline. The repository also has serving and on-device kits.

Title: What the repository holds
Tool | What it does
pipeline or AutoModel | Runs the model in plain PyTorch, with an ONNX copy of the same network
classify_long_audio | Finds speech with Silero VAD and returns a language timeline and shares
triton_inference | Serving kit for NVIDIA Triton, from one GPU to many nodes, with a CPU-only path
mobile_inference | On-device kit built on sherpa-onnx for Android, iOS and ARM Linux
notebook.ipynb | A Colab quick start that runs every example

Example: You pass in a short recording of someone speaking Hindi. The model returns the language label with a score you can treat as a confidence. For a long call with two languages, the long-audio helper returns a timeline and the share of speech in each.

For newcomers and experts: If you are new, language identification is the step that answers which language a clip is in, before anything else happens. A first try is the Colab notebook, which runs the card's examples.

For experts, the scores listed are 0.914 macro-F1 averaged across 37 test conditions, and 0.897 over all 23 languages. The card says no other system on that test goes above 0.713. Expected calibration error is 0.008 once one temperature setting is applied. Music clips and non-Indian speech get turned away, with AUROC from 0.89 to 0.91.

Audio clip goes in
A Conformer encodes the speech
Calibrated scores for 23 labels
Top language or a timeline comes out

Good to know:

• Scope: 22 scheduled Indian languages and Indian English, 23 labels in total.
• Short clips: the card claims a large lead at one second of audio.
• License: Apache 2.0 for both code and model.
• Cost to run: none named.
• Data it reaches: the audio you pass in, plus downloads from Hugging Face. The on-device kit runs with no server.
• Not for: identifying languages outside this set, or reading music as speech.

Title: Macro-F1 across all 23 languages
Anuvani-LID: 0.897 F1
Best other system, at most: 0.713 F1

On Hugging Face, 13 November 2026

• Task: audio classification
• Library: transformers
• Parameters: 116M
• Weights: about 463 MB in F32; about 58 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 276
• Likes: 7
• License: apache-2.0
• Languages: as, bn, brx, doi, en, gu, hi, kn
• Tags: transformers, onnx, nemo, safetensors, anuvani_lid, audio-classification, audio, speech, language-identification, spoken-language-id, indic, conformer

Links

• Model: https://huggingface.co/CIIL/Anuvani-LID
• Files: https://huggingface.co/CIIL/Anuvani-LID/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download CIIL/Anuvani-LID

Then follow the usage in its model card. Nothing in this kit runs by itself.