⚡ Vaani1-Lite-ASR-Turbo (Multilingual)
High-Speed Production Engine by Sorika Labs
CTranslate2 (faster-whisper) format of sorika-labs/vaani1-lite-asr
Vaani1-Lite-ASR-Turbo is the official CTranslate2-optimized release of the Vaani1-Lite-ASR model. It provides 4x-5x faster inference, cuts memory consumption by 50% (only ~484MB), and delivers real-time speech-to-text with latency under 150ms across 99+ languages (Hindi, English, Chinese, Japanese, Korean, Spanish, etc.).
🌍 Multilingual Capabilities
- Hindi (
hi): SOTA fine-tuned accuracy (12.68% WER, beats Whisper-Large-v3). - English (
en): 3.70% WER (Zero catastrophic forgetting). - East Asian (
zh,ja,ko): Full transcription support for Chinese, Japanese, and Korean. - Indic Languages: Bengali, Tamil, Telugu, Marathi, Gujarati, Urdu, Kannada, Malayalam, Punjabi.
- Auto-Detection: Automatically identifies the spoken language from audio.
🚀 Quickstart: Real-Time Transcription
Install faster-whisper:
pip install faster-whisper
Transcribe with Python:
from faster_whisper import WhisperModel
# Load the turbo model on CPU (uses ~250MB RAM)
model = WhisperModel("sorika-labs/vaani1-lite-asr-turbo", device="cpu", compute_type="int8")
# Transcribe Hindi, English, Chinese, Korean, or Auto-Detect:
segments, info = model.transcribe("sample_audio.wav", vad_filter=True)
print(f"Detected Language: {info.language} ({info.language_probability:.2f})")
for segment in segments:
print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")
🔗 Models in this Family
| Model | Format | Best For | Link |
|---|---|---|---|
Vaani1-Lite-ASR | PyTorch / Transformers | Fine-tuning, PyTorch pipelines, Colab demo | Hugging Face |
Vaani1-Lite-ASR-Turbo | CTranslate2 / INT8 / FP16 | Production APIs, Web Apps, Edge & Real-time bots | Hugging Face |
📜 License & Developer
- License: MIT (Permissive open source)
- Developer: Sorika Labs