---
title: Varuna STT
emoji: 🌊
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: "6.3.0"
python_version: "3.10"
app_file: app.py
pinned: false
license: other
license_name: skunkworks-modified-mit
short_description: Hindi ASR — upload mono/stereo, transcript per-channel
models:
  - SkunkWorkLabs/varuna-stt
  - nvidia/nemotron-speech-streaming-en-0.6b
datasets:
  - SkunkWorkLabs/hindi-asr-benchmark
hf_oauth: false
hardware: zero-gpu
---

# Varuna STT 🌊

Hindi ASR by **SkunkWorks Labs** — a 0.6 B-parameter Conformer-RNNT model
fine-tuned from NVIDIA's
[`nemotron-speech-streaming-en-0.6b`](https://huggingface.co/nvidia/nemotron-speech-streaming-en-0.6b).
Upload mono or stereo audio; stereo channels are transcribed independently
(useful for call-center recordings where the agent and customer are on
separate channels).

> ⚡ **This Space runs on ZeroGPU (NVIDIA A100).** Inference is fast —
> a 5-second clip transcribes in under a second. The same weights hit
> **RTFx 25×** (≈175 ms p50 latency per clip) on an H100; quality is
> identical across hardware.

## What is "Nemotron 0.6 B streaming, cache-aware"?

The base model is a **streaming, cache-aware Conformer**. Two things to know:

- **Streaming** — the encoder consumes audio in small chunks (tens of ms)
  rather than requiring the full utterance up-front. This means you can
  start emitting tokens while the speaker is still talking, instead of
  waiting for end-of-speech.
- **Cache-aware** — attention key/value states are carried over across
  chunks. The model "remembers" what came before without re-encoding the
  whole audio every step, so per-chunk compute is bounded and constant in
  memory.

**Why this matters for voice calling / IVR / real-time agents:**

- **Low first-token latency** — partial transcripts within a few hundred ms
  of speech start, suitable for live captions and barge-in detection.
- **Bounded compute per chunk** — predictable latency under load; a single
  GPU can serve many concurrent calls because per-stream memory stays flat.
- **No "wait until they stop talking"** — important for natural turn-taking
  in voicebots; the LLM downstream can start reasoning while the user is
  still speaking.
- **Per-channel stereo** — agent/customer separation comes for free when
  call audio is recorded on two channels (this Space demos that).

## Why Varuna

- 🥇 **Best WER on indictts (9.75 %)** out of Varuna, ElevenLabs Scribe v1,
  Deepgram Nova-2, and Sarvam Saarika v2.5 on the
  [SkunkWorkLabs Hindi ASR benchmark](https://huggingface.co/datasets/SkunkWorkLabs/hindi-asr-benchmark).
- **Trained on ~3 K hours of Hindi** (Shrutilipi, IndicVoices, Kathbath,
  Gramvaani, Vaani, Lahaja, IndicTTS, …).
- **ITN output by default** — digits, ordinals (`1st`/`3rd`),
  Indian-numbering commas (`2,50,000`), and Devanagari punctuation (`।`).
  Drops straight into IVR / transcription pipelines with no separate ITN
  postprocessor.
- See the [model card](https://huggingface.co/SkunkWorkLabs/varuna-stt) for
  full per-subset WER/CER tables, training mix, and inference example.

## Limitations

- **No code-switching yet.** Trained on monolingual Hindi audio.
  Hindi-English mid-sentence switches may produce transliteration
  artifacts. Bilingual fine-tune is on the roadmap.
- **Codec-degraded audio** (telephony, heavy compression) is weaker than
  studio-clean speech.
- **16 kHz mono internally** — other rates are resampled automatically.

## 📬 Contact

Need help with the **training recipe** or want to **fine-tune Varuna** on
your own data? Reach out: **harshris2314@gmail.com**.