More screenshots
Works with the AI Agents you already use
NOVA Hinglish STT
Speech-to-text for English and Hinglish that outputs Romanized Hindi instead of Devanagari using faster-whisper.
$5
NOVA Hinglish STT
Example session with this skill installed
I need a Python script to transcribe 'audio_input.wav' which contains a mix of Hindi and English. It must output the Hindi parts in Roman script.
- Read your context and instructions
- Compiled the nova hinglish stt
I have configured the transcription script using the Nova Whisper Hinglish model. This setup includes Voice Activity Detection (VAD) to handle silence and uses float16 compute for GPU efficiency.
Hinglish Transcription Script
import os
import sys
from faster_whisper import WhisperModel
# Ensure CUDA DLLs are in path for Windows GPU usage
# os.environ["PATH"] = r".venv\Lib\site-packages\nvidia\cublas\bin;" + os.environ["PATH"]
def transcribe_hinglish(audio_path):
# Initialize model on GPU with float16 for speed
model = WhisperModel(
"AyushM1701/nova-whisper-hinglish",
device="cuda",
compute_type="float16"
)
# Custom VAD to avoid clipping soft word endings
vad_params = {
"threshold": 0.35,
"speech_pad_ms": 300,
"min_silence_duration_ms": 500
}
# language="hi" forces Romanized Hindi output
segments, info = model.transcribe(
audio_path,
language="hi",
vad_filter=True,
vad_parameters=vad_params,
beam_size=1,
condition_on_previous_text=False
)
return " ".join(s.text.strip() for s in segments)
if __name__ == "__main__":
result = transcribe_hinglish("audio_input.wav")
print(f"Transcription: {result}")
Next steps
- Install dependencies using
pip install faster-whisper "av<16". - If running on Windows with a GPU, verify the CUDA 12 bin folders are added to your environment
PATH. - Run the script; the first execution will download the 1.6 GB model weights.
Connects securely to your tools. The creator never sees your data.
About this skill
Standard Whisper models fail on Hinglish, outputting messy Devanagari or hallucinating translations. This skill implements Nova Whisper Hinglish to provide transcription that writes Hindi in Roman letters, exactly as users type it.
What it does
- Romanized Hindi output ensures "kal meeting cancel kar do" stays in Roman script rather than Devanagari.
- Multilingual switching detects English or Hindi for each clip, so users can switch languages from one sentence to the next.
- Optimized performance uses CTranslate2 via
faster-whisperfor high-speed inference on CPU or GPU. - Voice activity detection applies custom VAD parameters to trim silence without clipping soft word endings.
- Offline execution runs entirely on local hardware after the initial model download.
How it works
- Environment setup creates a virtual environment and pins specific library versions like
av<16to prevent runtime crashes. - Model initialization loads the 1.6 GB
nova-whisper-hinglishweights with optimized compute types likeint8orfloat16. - Language detection analyzes audio clips to choose between
enorhimodes before processing. - Transcription processes an audio file or array with the silence filter on and condition_on_previous_text=False, so one bad segment can’t repeat through the rest.
Frameworks & tools
This skill uses Python, faster-whisper, CTranslate2, and NVIDIA CUDA 12 libraries. It specifically utilizes the AyushM1701/nova-whisper-hinglish model from Hugging Face.
Why this beats prompting it yourself: Prompting an LLM to romanize a Devanagari transcript adds latency and token costs. This skill fixes it in the model weights: on the MUCS 2021 Hinglish test set it makes about 40% fewer word errors than stock Whisper large-v3-turbo (45.7% vs 77.9%, after converting stock Whisper’s Devanagari output to Roman letters), at the same speed.
Use cases
- Message dictation for apps where users prefer typing Hindi in the Roman alphabet.
- Customer support analysis for Indian markets where Hinglish is the primary spoken dialect.
- Meeting transcription for bilingual teams switching frequently between English and Hindi.
- Voice commands for IoT devices or local bots in the Indian subcontinent.
Known limitations
Needs about 2 GB of disk space and Python 3.9+. CPU only on Mac, so expect about a second or more per clip. On a Windows GPU, the CUDA bin folders must be on the PATH of the terminal that runs it. The model is licensed CC BY-SA 4.0. The Hinglish test clips come from the same corpus as most of the training data, so expect a smaller gain on other Hinglish.
Requires 2GB disk space. Python 3.9+ required. Windows GPU usage requires manual PATH configuration for CUDA DLLs. Only supports Romanized script for Hindi, not Devanagari.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 5 days ago