⚡ Vaani-2 Flash ASR

High-Throughput, Low-Latency Speech Foundation Model for Production APIs & Real-Time Streaming

Developed by Sorika Labs

⚡ Try ZeroGPU Live Demo • Hugging Face Model • Sorika Labs Organization

---

⚡ Try the Live ZeroGPU Demo

You can test Vaani-2 Flash in real-time with zero installation directly in your browser:

👉 Launch Vaani-2 Flash Web UI & API Demo (NVIDIA A10G GPU)

• 🎙️ Microphone & File Uploads: Test .mp3, .wav, .m4a, .flac, .aac recordings instantly.
• ⚡ Sub-Second GPU Speed: Real-time < 0.8s latency powered by Hugging Face ZeroGPU (NVIDIA A10G).
• 🌐 Auto & Explicit Modes: Automatic speech detection or dedicated Hindi / English transcription.

---

🌟 Overview & Key Breakthroughs

Vaani-2 Flash is Sorika Labs' flagship high-throughput speech recognition foundation model. Built on top of OpenAI's 809M-parameter whisper-large-v3-turbo, Vaani-2 Flash delivers an unmatched combination of extreme inference velocity (up to 30.3x Real-Time) and flawless English and Indian recognition accuracy.

In our comprehensive 10-benchmark evaluation, Vaani-2 Flash achieved the #1 lowest English Average WER (5.90%) across all evaluated models, beating both Alibaba's Qwen3-ASR (6.19%) and OpenAI Whisper baselines (8.05%).

Key Highlights:

• 🏆 #1 Overall English Accuracy (5.90% Avg WER):
  • VoxPopuli: 2.22% WER (Rank 1)
  • SPGISpeech: 3.55% WER (Rank 1)
  • Earnings22: 8.62% WER (Rank 1)
  • LibriSpeech Clean: 4.84% WER
• ⚡ Blazing Inference Velocity (21.5x Real-Time): Thanks to the optimized 4-layer decoder of Whisper-Turbo, Vaani-2 Flash processes 30 seconds of audio in under 1.0 second on GPU, making it ideal for live streaming and enterprise APIs.
• 🇮🇳 Robust Indian Speech Recognition: Achieves 16.38% on FLEURS Hindi and 14.44% on Private Conversational Hindi, slashing Whisper-Small's 47.8% error rate by more than 60%.
• 🛡️ Zero Catastrophic Forgetting: Preserves global multilingual capabilities across European and Asian languages without weight drift.

---

🧭 Sorika Labs Vaani Speech Series Portfolio

The Vaani family is engineered by Sorika Labs to provide production-ready speech intelligence across diverse computational constraints:

---

📊 Grand Benchmark Leaderboard (10 Standardized Benchmarks)

All models were evaluated under identical real-world conditions across 7 English benchmarks (meeting, financial, podcast, clean read, noisy speech, parliamentary) and 3 Indic/Hindi benchmarks:

Lower WER indicates superior recognition accuracy. Speed is measured in Real-Time Factor (RTFx) on NVIDIA GPU.

---

🚀 Quickstart & Inference

1. Using Hugging Face pipeline

import torch
from transformers import pipeline

device = "cuda:0" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if torch.cuda.is_available() else torch.float32

pipe = pipeline(
    "automatic-speech-recognition",
    model="sorika-labs/vaani2-flash",
    torch_dtype=dtype,
    device=device,
    chunk_length_s=30,
    trust_remote_code=True
)

# 1. Transcribe English Audio (Ultra-Accurate 5.90% Avg WER)
result_en = pipe("english_call.wav", generate_kwargs={"language": "en", "task": "transcribe"})
print("English:", result_en["text"])

# 2. Transcribe Hindi Audio
result_hi = pipe("hindi_audio.wav", generate_kwargs={"language": "hi", "task": "transcribe"})
print("Hindi:", result_hi["text"])

# 3. Automatic Language Detection & Transcription
result_auto = pipe("mixed_speech.mp3", generate_kwargs={"task": "transcribe"})
print("Transcription:", result_auto["text"])

2. Using AutoModelForSpeechSeq2Seq & AutoProcessor

import torch
import torchaudio
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor

model_id = "sorika-labs/vaani2-flash"
device = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if torch.cuda.is_available() else torch.float32

processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id,
    torch_dtype=dtype,
    low_cpu_mem_usage=True,
    trust_remote_code=True
).to(device)

# Load 16 kHz Mono Audio
audio, sr = torchaudio.load("sample.wav")
if sr != 16000:
    audio = torchaudio.transforms.Resample(sr, 16000)(audio)

inputs = processor(audio.squeeze().numpy(), sampling_rate=16000, return_tensors="pt")
input_features = inputs.input_features.to(device, dtype=dtype)

with torch.no_grad():
    predicted_ids = model.generate(input_features, language="hi", task="transcribe")

transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print("Transcription:", transcription[0])

---

🔮 Roadmap: Vaani-2.1 Flash

Sorika Labs is actively fine-tuning the next revision, vaani2.1-flash, which pairs the blazing 25x+ RTF Whisper Turbo architecture with a high-fidelity curated Hindi & Hinglish conversational corpus. Target goals:

• Sub-10% Hindi Average WER
• Native Hinglish code-switching normalization
• Retaining the #1 English benchmark ranking

---

⚙️ Architecture & Model Specs

• Base Model: openai/whisper-large-v3-turbo (809 Million Parameters)
• Audio Encoder: 32-layer Transformer, 1280 hidden size, 128 mel bins
• Decoder: 4-layer fast Transformer decoder (4x faster than standard large-v3)
• Weights: 100% permanently fused model.safetensors (1.62 GB FP16)
• License: Apache 2.0

---

🏢 Developed by Sorika Labs

Sorika Labs — Advancing open-source speech AI for Indian and global enterprises.