Indri: Ultralight Audio Model

Overview
Indri is a novel, ultra-small transformer-based TTS/ASR model that treats audio as discrete tokens. It supports both Hindi and English, and delivers high-quality, style-consistent speech synthesis and recognition—all in a 124 M-parameter footprint.

• HF : https://huggingface.co/11mlabs/indri-0.1-124m-tts
• Blog : https://www.indrivoice.ai/blog

Key Features

• Tiny & Fast: Just 124 M parameters (GPT-2 small backbone) yet achieves realtime on CPU and up to 10× realtime on consumer GPUs (e.g. RTX6000 Ada: 400 tokens/s, <20 ms to first token).
• Streaming & Voice-Cloning: Autoregressive audio-token decoding with streaming output; supports speaker style prompts (<5 s) for consistent voice cloning.
• Bilingual & Code-Mixing: Natively handles English, Hindi, and mixed-language inputs.
• End-to-End Pipeline:
  1. Text → text-tokens
  2. GPT-2 LM → audio-tokens
  3. Mimi decoder → waveform
• Novel architecture combining transformers with audio tokens, bringing text and audio to same space.

Audio Samples

• Sample 1: मित्रों, हम आज एक नया छोटा और शक्तिशाली मॉडल…
• Sample 2: भाइयों और बहनों, ये हमारा सौभाग्य है कि…
• Sample 3: Hello दोस्तों, future of speech technology mein…
• Sample 4: In this model zoo, a new model called Indri…

Links & Resources

• Blog post: Building Indri TTS
• Model & demos: 11mlabs/indri-0.1-124m-tts on Hugging Face
• Code & self-hosted service: github.com/indri-voice/indri