SonicLM
SonicLM: Open‑source, real‑time speech AI for human‑like voice interactions
Quick facts
- Best for
- SonicLM: Open‑source, real‑time speech AI for human‑like voice interactions
- Pricing
- Freemium
- Editor rating
- 4.5 / 5
- Community saves
- 0
About SonicLM
SonicLM is an open-source suite of speech foundation models for real-time, human‑like voice interactions. Built for low-latency performance, SonicLM delivers sub‑200ms streaming ASR, speech‑to‑speech translation, and audio understanding without relying on slow text intermediaries. Trained on 1M+ hours of multilingual audio, it supports English, Spanish, German, French, Hindi, and more with strong zero‑shot generalization. The models run efficiently on consumer hardware (Apple Silicon via MLX, as well as NVIDIA GPUs), and ship with model cards, demos, and benchmarks on Hugging Face. With Apache 2.0 licensing, SonicLM empowers developers and researchers to build voice agents, live captioning, and interactive AI experiences at state‑of‑the‑art quality and speed.
Pros
- Sub-200ms real-time latency on consumer hardware
- Multilingual ASR and speech-to-speech translation with zero-shot generalization
- Open-source Apache 2.0 license with models and code on Hugging Face
- Optimized inference via MLX for Apple Silicon; supports NVIDIA GPUs
- Joint speech–text token modeling enabling rich audio understanding
- Streamed, low-latency decoding 3–6x faster than common baselines like Whisper
- Human-like speech generation and editing with voice, timing, and prosody preservation
- Model sizes for different budgets (e.g., compact 1.5B and high-performance 8B)
- State-of-the-art results on Speech Arena, CoVoST2, and FLEURS benchmarks
- Live demos, model cards, and installation guides for quick developer onboarding