Skip to content
Audio

Vaani-3 ASR

compact speech recognition with emotion and intent analysis

2 min read

This model from sorika-labs transcribes spoken audio into clean, formatted text while simultaneously detecting the speaker's emotions, tone, sentiment, intent, and urgency in a single pass. Built on a 0.46B-parameter hybrid architecture, it pairs the acoustic encoder (audio processing block) from Qwen/Qwen3-ASR-1.7B with the google/gemma-3-270m text decoder and a dedicated classification module. Unlike standard ASR models, it adds punctuation out of the box, formats numbers, dates, and currencies (ITN — inverse text normalization), and handles mixed-language speech without requiring a language tag upfront. It is built for developers working on voice bots, call centers, and edge devices supporting English and 12 Indian languages.

Key features

  • Dual output in a single pass: alongside the transcript, it returns structured JSON scoring tone, emotions, sentiment, call intent, and urgency.
  • Ultra-compact and fast: its 0.46B-parameter footprint delivers speeds ranging from 8.1x real-time on a Raspberry Pi 5 up to 62.1x on an NVIDIA A100.
  • Support for 12 Indian languages and English: reliably transcribes spontaneous speech and code-switching (switching between Hindi and English mid-sentence).
  • Ready-to-use quantized builds: available in GGUF (starting at 283.5 MB for Q4_0) and ONNX INT8 (150 MB) formats for low-end hardware and mobile apps.

What you can do

  • Automate call center analytics: transcribe call recordings and immediately route urgent or negative tickets without running a second NLP model.
  • Embed into offline assistants: run the Q4_0 or ONNX build on laptops, smartphones, or single-board computers like the Raspberry Pi 5 using roughly 300–460 MB of memory.
  • Process long audio recordings: native generation length is extended to 4,096+ tokens (configurable via max_new_tokens), making it practical for long voice clips.

Example

import torch
from transformers import AutoModel

device = "cuda" if torch.cuda.is_available() else "cpu"
model = AutoModel.from_pretrained("sorika-labs/vaani-3-asr", trust_remote_code=True, torch_dtype=torch.float32).to(device).eval()
transcript = model.transcribe("sample.wav", language="hi", max_new_tokens=4096)
print("Transcript:", transcript)
Read moreDetails
Quality metrics (WER — word error rate, lower is better): AI4Bharat Kathbath (Hindi) — 3.4%, AI4Bharat Lahaja (Hindi) — 8.8%, 12-language average on Google FLEURS — 12.7%.
Latency on 10 seconds of audio: 161 ms on NVIDIA A100 (BF16), 198 ms on RTX 4090 (FP16), 510 ms on Apple M2/M3 (GGUF Q4_0), 1,240 ms on Raspberry Pi 5.
Supported languages: English, Hindi, Bengali, Marathi, Telugu, Tamil, Gujarati, Kannada, Malayalam, Punjabi, Assamese, Odia.
License: Apache 2.0 and Gemma Terms of Use.
Discuss Telegram 0 2

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # sorika-labs/vaani-3-asr
    # ===== Tokens: paste yours between the quotes =====
    export HF_TOKEN=""        # huggingface.co/settings/tokens, for gated models
    # ==========================================
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download sorika-labs/vaani-3-asr --local-dir ~/models/vaani-3-asr
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to vaani-3-asr
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of vaani-3-asr (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

sorika-labs/vaani-3-asr
Author
sorika-labs
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

20

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments

Computer vision

Gemma 4 E4B

Gemma 4 E4B