Introducing Vaani 1 Lite: Streaming Multilingual Speech Recognition for Indic Dialects
- Sub-100M parameter conformer-transducer streaming speech backbone
- Real-time streaming transcription across 22 scheduled Indian languages and mixed code-switching
- Robust acoustic noise cancellation with sub-100ms time-to-first-token audio latency
Today, Sorika Labs is announcing Vaani 1 Lite, our streaming multilingual speech recognition engine engineered for high-accuracy transcription across low-resource Indic dialects.
The Challenge with Low-Resource Indic Speech
Overcoming acoustic diversity and mixed code-switching.
Traditional ASR engines fail dramatically in real-world Indian acoustic environments due to heavy background noise, regional tonal inflections, and spontaneous code-switching between Hindi, regional languages, and English.
Vaani 1 Lite solves this by training a compact conformer-transducer architecture directly on multi-dialect acoustic audio datasets, maintaining sub-100ms end-to-end streaming latency without cloud reliance.
Developer API Integration
Streaming transcription in minimal lines of code.
Developers can now query the Vaani 1 Lite engine through our low-latency streaming WebSocket endpoint or run the quantized model locally on edge silicon:
import sorika
client = sorika.Client(api_key="sk_live_...")
stream = client.speech.transcribe_stream(
audio_source="mic_stream",
model="vaani-1-lite-asr",
languages=["hi", "en", "ta", "te"]
)
for chunk in stream:
print(chunk.text, end="", flush=True)