Description
Audio8 ASR Infinite is a cutting-edge streaming speech recognition model developed to provide highly responsive and accurate transcription services. Built on a native streaming architecture, it decodes audio at a rate of 12.5 times per second, ensuring minimal latency and high efficiency. The model supports unlimited-length transcription, made possible by a rolling KV cache that maintains constant memory and latency, even during 24/7 operation.
Read More
One of the standout features of Audio8 ASR Infinite is its selectable audio clock, which allows users to choose between 80, 120, or 160 ms settings. This flexibility helps balance perception granularity with resource cost. Additionally, the model offers configurable transcription delay options ranging from 240 to 560 ms, enabling users to trade off delay for improved accuracy.
The model is bilingual, supporting both Chinese and English languages, and incorporates semantic voice activity detection (VAD) to distinguish between thinking pauses, stuttering, and actual end-of-turn speech, where traditional acoustic VAD often fails. This feature enhances the model's ability to provide more natural and accurate transcriptions.
Audio8 ASR Infinite inherits the Voxtral real-time audio architecture and DSM-style streaming, featuring components such as the Causal Audio Tower and Audio Projector. The model's architecture includes 32 layers in the audio tower and 36 layers in the text decoder, with a vocabulary size of 151,936.
Currently in its preview release stage, Audio8 ASR Infinite delivers a robust transcription base, with future updates planned to include frame-level semantic perception capabilities. The model has been evaluated against various test sets, demonstrating competitive performance in terms of character error rate (CER) and word error rate (WER) compared to other models.
Designed for developers and businesses seeking reliable and scalable speech recognition solutions, Audio8 ASR Infinite offers programmatic simulated-streaming decode capabilities and supports deployment via Docker compose, making it accessible for integration into various applications.
- CategoryAI Transcription Tools
- Pricing ModelOpen source
- DeploymentCloud
- PlatformsWeb
Audio8 ASR Infinite Highlights
Native streaming architecture
Unlimited-length transcription
Selectable audio clock settings
Configurable transcription delay
Bilingual support: Chinese and English
Semantic voice activity detection
Voxtral real-time audio architecture
Rolling KV cache for constant latency
Getting Started with Audio8 ASR Infinite
Access model: Visit the Hugging Face model page
Authenticate: Set up necessary credentials
Set up environment: Prepare your development environment
Integrate via API: Use the provided API for integration
Optimise: Adjust settings for optimal performance
Audio8 ASR Infinite's Use Cases
- Real-time Transcription
- Bilingual Transcription
- 24/7 Monitoring
- Semantic Analysis
- Customizable Delay
