Question 1

How do I know what is causing my voice agent's failures?

Accepted Answer

If you want to understand your voice agent's failures, look at the audio itself, not just the transcript. A model (like Tyto) that scores each call across specific quality dimensions, like background noise, reverb, interfering speech, or packet loss, will tell you which one is actually the driver, instead of leaving you to guess from a text-only transcript or a frustrated caller.

Question 2

Why does my voice agent work in testing but fail on real calls?

Accepted Answer

Test audio is almost always cleaner than production audio. Real calls bring bad microphones, phone-network compression, background noise, multiple speakers, and other conditions a demo environment doesn't reproduce.

Question 3

What is ai-coustics and what problem does it solve for voice AI?

Accepted Answer

ai-coustics is the audio intelligence layer for Voice AI. Founded in Berlin in 2021, we're a team of audio natives building the SDK that enhances, isolates, and diagnoses speech in real time, so voice agents get clean, reliable audio input before it reaches ASR and your LLM.

Question 4

What is Word Error Rate and why does it matter for voice agents?

Accepted Answer

WER measures how often a transcript is wrong by counting substitutions, deletions, and insertions versus a reference transcript. For voice agents, WER is one of the main evaluation metrics and it matters because transcription errors cause issues downstream. Bad STT leads to wrong intent detection, confused LLM responses, failed tool calls, and awkward turn-taking.

Question 5

What is the difference between real-time and post-call monitoring?

Accepted Answer

Real-time (streaming) monitoring scores audio as the call happens, so you can catch or flag a problem while the call is still live. Post-call (batch) monitoring scores recordings after the fact, which is better suited to diagnosing patterns across hundreds or thousands of calls at once, like finding out a specific carrier route is causing packet loss.

Question 6

How can I monitor voice agent call audio quality without listening to every call?

Accepted Answer

Run your calls through an audio-scoring model that's cheap enough to use on every recording instead of manual spot-checks. Tyto, ai-coustics' audio insight model, processes an entire archive of calls and returns a risk score for each one, so you review the calls it flags instead of sampling a handful and hoping they're representative.

Question 7

What is Tyto and what does it do?

Accepted Answer

Tyto is ai-coustics' audio insight model: it scores voice-agent call audio and predicts, per 5-second window, how likely that audio is to break downstream steps like voice activity detection, turn-taking, or speech-to-text. It runs on-prem, on CPU, and is cheap enough to run on every call rather than a sample.

Question 8

How does Tyto's risk score work?

Accepted Answer

Every call gets a single Risk Score plus a breakdown across quality dimensions (including background noise, reverb, interfering speech, packet loss, and codec degradation) so you get both a headline number and the reason behind it. Scores generally fall into three bands: Good (below 0.35), Warn (0.35–0.60), and Bad (above 0.60).

Question 9

Is my call audio data private if I use ai-coustics Tyto?

Accepted Answer

Yes. Tyto runs entirely within your own infrastructure, and audio is processed on-device. No audio data is sent to ai-coustics' servers. Only usage telemetry (minutes processed) is sent back for billing.

Question 10

What should I do when Tyto flags a call?

Accepted Answer

Check which dimension is driving the risk score, since it points directly to the fix: high background noise or interfering speech generally calls for a speech-enhancement model or a specialised VAD, high reverb is often a mic-placement issue worth flagging for review, and high packet loss usually points to a network or infrastructure problem rather than an audio-model fix.

Audio Insight into every call for Voice AI teams

Audio Insight into every call for Voice AI teams

One score, all your models

Six dimensions of audio insight

Score every call, on-premise

Tyto

Audio Insight

Try it yourself

Real-time

Post-call

Your questions, answered

Bring real-time audio intelligence into your voice AI stack

Bring real-time audio intelligence into your voice AI stack

Bring real-time audio intelligence into your voice AI stack

Quail

Sparrow