Tyto is a real-time audio quality diagnostic model released by ai-coustics GmbH in June 2026. It pre-analyzes the audio state at the input stage of a speech AI pipeline to predict the potential for malfunctions in downstream models (STT, VAD, turn-taking, speech-to-speech conversion). Similar to how a spell checker is used before feeding text to GPT, Tyto can be considered the "stethoscope for speech AI," quantifying how problematic an audio input is with a Tyto Risk Score ranging from 0 to 1 before the speech AI processes it. It operates with a latency of less than 30ms (based on 16kHz PCM) using only the CPU, and does not require a GPU, allowing it to be integrated into existing speech pipelines without additional hardware investment. Existing speech AI systems lacked visibility into input audio quality. When noisy or packet-loss-ridden calls came in, STT would misinterpret, VAD would mistake background noise for speech, and turn-taking would interrupt at inappropriate times, leading to a chain of errors. However, it was only discovered afterward that the root cause was not the model itself, but the input audio. Tyto solves this problem at the input stage. It measures six quality dimensions – Noise (environmental noise), Speaker Reverb, Speaker Loudness, Interfering Speech, Background Media Speech, and Packet Loss – in real-time on a scale of 0 to 1, and calculates a composite Risk Score. If the Risk Score is below 0.35, it is classified as safe (Green); if it is between 0.35 and 0.60, it is classified as a warning (Warn); and if it is above 0.60, it is classified as a high probability of severe downstream errors (Bad). In real-world speech AI operating environments, Tyto can be used in three main ways. In real-time mode, it calculates scores in 5-second windows, allowing the agent turn detection to be switched to a more conservative setting when noise spikes, or injecting the context "current call audio quality is poor" into the LLM prompt to improve response quality. In offline mode, it performs batch analysis on 100% of the entire call to triage the worst calls based on the p95 Risk Score and identify the primary cause of quality degradation using the argmax for each dimension. Thirdly, the Aware-Tuned-Reactive three-stage adaptive mode, demonstrated in ai-coustics' demo, is a pattern that automatically switches agent behavior based on the Risk Score level, systematically increasing the noise resilience of the speech agent.