TL;DR

• Tyto is a lightweight ai-coustics model that predicts whether audio is likely to cause a failure in a Voice AI pipeline before it happens.

• It can run in two modes. Real time, so the system can react mid-call, or post-call for batch analysis.

• It scores audio across six dimensions: Noise, Interfering Speech, and Speaker Reverb affect whether the speaker can be understood; Packet Loss and Codec Degradation are technical failures; Speaker Loudness is informational only.

• The resulting Risk Score predicts the likelihood of agent failure.

• Available now in the ai-coustics SDK, free to test.

Voice assistants, transcription tools, and customer service bots are usually trained and tested on clean audio. Real calls are not clean. People call from noisy kitchens, bad phone lines, echoey rooms and cheap headsets. When the audio is rough enough, these systems get it wrong: they mishear a word, cut someone off or fail to respond. Often nobody notices until a customer complains.

Tyto: Audio insight for your voice interface

Tyto is a lightweight model built by ai-coustics that listens to incoming audio and predicts whether it is likely to cause a failure in a Voice AI pipeline. Rather than cleaning up the audio, it evaluates it and reports whether it can be trusted. Its main output is a Risk Score between 0 and 1, low means the audio is fine, high means a failure is likely.

The Risk Score is the warning, flagging a call that is likely to fail before it happens. The six dimensions give you the why: Noise, Interfering Speech, Speaker Reverb, Packet Loss, Codec Degradation, and Speaker Loudness.

Beyond the single Risk Score, Tyto gives your team a way to react to changing audio conditions as they happen. Tyto is built for machine understanding but by reacting in real time, that can look like a human reaction - such as falling back to manual turn-taking or prompting a caller to move somewhere quieter. It also tells you which of the six dimensions is degrading a call, and how that's likely to affect the models reading it downstream. Your team can diagnose audio performance without manual listening or guessing from a transcript.

The dimensions that affect intelligibility

Here are the three dimensions that affect the intelligibility of processed voice in the context of machine understanding. These are likely to cause Voice Activity Detection (VAD) misfires and transcription errors.

Noise measures how loud the speaker’s background is, whether that is a running vacuum, traffic or loud music. Since ASR models are already quite resistant to noise, it is not simply about whether background sound is present. This dimension helps you identify calls with loud or intense noise, which, when paired with the Risk Score, allows you to understand the cause of downstream failure. For example, a caller phoning in from a moving car with their window down, engine and road noise included, is an example of the speech being present, but buried under sound loud enough to cause a downstream transcription errors.

Interfering Speech measures other voices competing with the main speaker, such as a second person talking, or a TV playing in the background. Both are the same underlying problem of another speech-like structure competing for attention. It is a very important metric to track if you only want your Voice AI application to react to the main speaker. A customer on speakerphone with a TV on is a typical case: even though the TV isn't a real conversational partner, its speech-like structure can pull turn-taking off track or cause overtranscription.

Speaker Reverb measures how much echo is on the speaker's voice. It is low when someone is close to the microphone and coming through clearly. It is higher when they are farther away, or in a space that reflects sound. In practice, that is the difference between a dry, close-mic sound and one that comes across as distant, reverberant, or hollow. For example, a call taken in an echoey hallway can result in words being dropped even when nothing else about the call is degraded.

Technical failure modes

The following two dimensions are responsible for detecting issues with the audio signal that can distort the human speech and cause unexpected behavior in your Voice AI application.

Packet Loss measures discontinuities in the audio stream: dropped frames, choppiness, clipped syllables. It typically comes from a weak network connection, or hardware that cannot keep up while processing the call. That includes a voice agent's own system falling behind and dropping frames in the audio it generates if you are using Tyto on the audio generated by your voice agent. For example, a caller on a weak mobile connection driving through an area with poor signal will show up here: syllables get clipped mid-word, which can cause the downstream model to change or drop words.

Codec Degradation flags the muffled, thin, or metallic sound that appears when audio is compressed too aggressively. It differs from packet loss: the audio stream stays continuous, but its quality has been reduced by compression. A common example is a codec like Opus running at a high compression ratio to save bandwidth. This also shows up on calls routed through carriers or older mobile networks that compress aggressively by default, even when the connection itself never drops a frame. Heavy compression can cause difficulties for VAD and turn-taking models.

Neutral dimension

Lastly, we also include one dimension that doesn’t drive pipeline failures but gives you more information about the acoustic conditions of the call.

Speaker Loudness reports a level rather than a problem: how loud the main speaker is. It is useful for catching a call that is far too quiet or far too loud, and unlike the other five dimensions, it is informational rather than a sign of failure. A caller who speaks close to the microphone will register a high loudness score without anything being wrong, while one speaking from across the room may register low enough that individual words become hard to pick up - this context can help you correctly understand the pipeline performance.

Measuring the score

Tyto scores audio in windows of adjustable length (for example 5 seconds on a rolling or sequential basis). In real time, each window's score is available as soon as it's produced, so your Voice AI agent or other application can react to conditions as they change during the call. When scoring recordings after the fact, those same per-window scores can be rolled up into per-call numbers instead, for example by taking the an average, a p95, or the percentage of the call spent in the Warn band or above - you can adjust that threshold yourself. This way you can rank a whole batch of calls by how much of each one was actually degraded, rather than by a single average that can hide a rough patch.

Noise, Interfering Speech and Speaker Loudness are all volume measurements. For these three, Tyto is built to be sensitive enough to but avoid being skewed by a single loud moment, such as a slammed door or one loud laugh, that does not reflect the rest of the call. It measures the fuller picture instead of reacting to a brief spike, so each score reflects the call as a whole rather than its loudest second.

Tyto dimensions summarized

Depending on your use case and call conditions, you can pay more attention to other dimensions.

Try out Tyto

Voice AI is already running in customer support, healthcare, and drive-thru ordering, among other industries, and Tyto gives your team insight into how well it's actually performing. It can be used in two ways. Across a batch of calls, the dimensions show which ones are noise problems and which are connection problems without listening to every recording. During a live call, a spike in noise or interfering speech can trigger the system to adapt, for example by prompting the user to take measures to reduce background noise.

Tyto is available in the ai-coustics SDK. Test it for free in our developer platform, or see the docs for integration details.