Audio8 ASR Infinite Targets 24/7 Transcription

Edge0 has released Audio8 ASR Infinite, an Apache 2.0-licensed streaming speech recognition model with 4 billion parameters. The company claims unlimited-duration transcription, sub-second latency, bounded memory use, and semantic end-of-turn detection. Those features target services that keep microphones open for hours, including meeting transcription, call analysis, and full-duplex voice agents.

Developers can select an 80, 120, or 160 millisecond audio clock and configure transcription delay from 240 to 560 milliseconds. One checkpoint therefore supports several latency budgets, allowing applications to favor faster responses or additional acoustic context at runtime.

The ten-hour state problem

Chunked streaming systems process short audio segments and join their outputs, while stateful systems retain context between steps. During long sessions, retained state can consume growing amounts of memory, and positional encodings can move beyond the ranges seen during training. Either problem can reduce accuracy or force periodic resets.

Audio8 limits its attention history with a rolling key-value (KV) cache. The cache stores 30 seconds of context, discards older entries, and applies exact re-basing to rotary positional embeddings (RoPE), which encode token positions inside attention. Re-basing returns those positions to the model’s familiar numerical range after each rollover.

• Context window: 30 seconds of rolling audio and text state.
• Position handling: Exact RoPE re-basing after old context leaves the cache.
• Decision rate: 12.5, 8.3, or 6.25 steps per second at the 80, 120, or 160 millisecond clocks.
• Output schedule: One text token can be emitted at each clock step.

Turn-taking beyond silence

Acoustic voice activity detection (VAD) usually estimates whether speech is present, making silence the main signal for turn boundaries. Audio8 adds semantic VAD heads designed to distinguish an actual end of turn from thinking pauses, hesitation, and stuttering. Voice agents can use those predictions to decide when to respond without relying on a fixed silence threshold.