OneASR Overview & Architecture
OneASR is an open-source, high-throughput speech recognition (ASR) aggregation gateway and self-hosted inference server. It standardizes diverse speech models into unified, OpenAI-compatible APIs and ultra-low latency WebSocket streams.
Why OneASR?
Modern AI applications and voice agents require reliable, low-latency, and private speech recognition. However, running state-of-the-art ASR models often introduces fragmented interfaces, missing VAD/chunking pipelines, and high cloud costs.
OneASR solves this by providing:
- Unified OpenAI Compatibility: Direct drop-in replacement for
POST /v1/audio/transcriptions. - Multi-Engine Multiplexing: Switch between Faster-Whisper, X-ASR (Zipformer), Qwen3-ASR, and FireRedASR with zero client code modifications.
- Integrated VAD & Chunking: Silero-VAD pipeline handles audio normalization, pause segmentation, and repetition filtering automatically.
- 100% Data Sovereignty: Self-hosted on your Mac, local GPU workstation, or Kubernetes cluster with zero external telemetry.
System Architecture
[ Clients & AI Apps ] (DuRT, Dify, Open WebUI, Python SDK, cURL)
│
▼ (HTTP / WebSocket)
┌────────────────────────────────────────────────────────┐
│ OneASR Gateway (FastAPI) │
│ │
│ ┌───────────────────────┐ ┌───────────────────────┐ │
│ │ Auth & API Layer │ │ ASR-Toolkit (VAD) │ │
│ │ • /v1/audio/* │ │ • Silero-VAD │ │
│ │ • /api/v1/stream │ │ • Dynamic Chunking │ │
│ └───────────┬───────────┘ └───────────┬───────────┘ │
│ │ │ │
│ ▼ ▼ │
│ ┌──────────────────────────────────────────────────┐ │
│ │ Engine Registry & Dispatcher │ │
│ └──────────────────────────┬───────────────────────┘ │
└─────────────────────────────┼──────────────────────────┘
│
┌────────────────────┼────────────────────┐
▼ ▼ ▼
┌──────────────────┐ ┌──────────────────┐ ┌──────────────────┐
│ Faster-Whisper │ │ X-ASR Zipformer │ │ Qwen3-ASR │
│ (CUDA/Metal/CPU) │ │ (ONNX / Stream) │ │ (vLLM / PyTorch) │
└──────────────────┘ └──────────────────┘ └──────────────────┘Core Capabilities
- Realtime Streaming: Sub-200ms chunk processing with WebSocket full-duplex protocol.
- Async Batch Jobs: Upload large multi-gigabyte audio/video files with background worker execution and progress polling.
- Millisecond Timestamps: Generate word-level and sentence-level timestamps for subtitle alignment and speaker diarization.
- Embedded Web Console: Interactive Vue 3 dashboard for microphone testing, model metrics, and task management.
Next Steps
- Docker Deployment Guide — Quickest way to launch OneASR with NVIDIA GPU support.
- Manual Installation — Install via Python virtualenv on macOS or Linux.
- Hardware Acceleration — Tune CUDA, Apple Silicon Metal, and CPU thread allocations.