Feature · Speech analytics
Speech analytics for what was said inside each call.
PyAI Hear writes down every word, then PyAI Recap tags the topics, decisions and promises buried in the conversation.
What PyAI gives you for speech analytics
- Speech analytics in PyAI pairs Hear transcripts, accurate to 4.76 percent on 8 kHz phone audio, with Recap tags at $0.02 a minute, so a 10 minute call costs 20 cents to read for topics and promises.
- Topic trackers learn a theme from a handful of example phrases, each one set to hear only the caller or only the agent, and every tagged insight opens the recording at the exact turn so nobody scrubs a 20 minute file.
- Batch jobs run 92 to 247 times faster than playback, so each archived hour finishes in under a minute with no platform fee.
What happens when a topic tracker fires?
Choose a conversation and press play. The lines are an example script, not a recording, and each tracker is one a support team would seed.
Pick a moment, then press play
How does speech analytics turn call audio into topics?
Speech analytics runs in two stages: transcription, then analysis of the text. PyAI runs both stages on the same call record, with Hear handling the words and Recap handling the meaning.
Recording lands
PyAI receives audio from a finished Omni conversation, a transcription jobs upload, or a JustCall call pushed through its private webhook.
Words become text
Hear cuts long audio into overlapping chunks, transcribes the chunks in parallel, and labels who spoke using voice differences or the stereo channel.
Meaning gets tagged
Recap marks decisions, promised follow-ups, topics and notable moments in the transcript, and can fill CRM fields you define.
Themes get counted
Each active tracker scans the fresh transcript. A manager then confirms or dismisses each hit, and every hit points to the exact turn.
Caller raises a tracked theme
The conversation is filed under that theme, and the hit opens the transcript at the caller's own sentence.
Agent promises a follow-up
Recap records the promise as an action item with its next step, and the Overview keeps it open until someone closes it.
Speaker labels come out reversed
A manager uses the role flip on the transcript tab, so trackers limited to caller speech stop reading the agent's lines.
Which questions does speech analytics answer for call centers?
Conversation analytics for call center teams checks every call, not just a few picked at random. Five teams use the spoken words most. Each asks its own question of the same transcripts.
Reasons people leave
A caller-only tracker seeded with lines like "thinking of switching" counts churn talk each week and shows whether price or service comes up more.
Requests in the customer's words
Recap topics cluster repeated asks from support conversations, and product managers play back the exact turn rather than reading a ticket summary.
Promises kept or dropped
The coach card plus open action items show whether an agent agreed to a next step before hanging up.
What changed after a launch
A tracker created after a price change also scans retained recordings, so the before and after counts use one definition.
Lines agents must never say
An agent-only tracker built from examples of forbidden guarantees checks new and retained conversations, so reviewers read matched turns instead of picking random files.
Speech analytics questions buyers ask
How is speech analytics different from call analytics?
Speech analytics studies what was said inside a conversation, such as topics, phrases and promises, whereas call analytics measures the call as an event, such as volume, duration, outcome and cost. PyAI covers the words with Hear and Recap, while the Analytics tab and AMD history cover call-level numbers.
Is ASR considered AI?
Yes. Automatic speech recognition uses machine learning models trained on recorded speech to convert audio into text, which makes ASR a form of AI. PyAI Hear is an ASR model built for 8 kHz phone audio and scored a 4.76% word error rate on real phone calls in PyAI's 24 June 2026 tests.
Does PyAI conversation analytics measure sentiment?
No. PyAI Recap does not output a sentiment or emotion score per conversation in current PyAI docs. Recap raises risk signals, open action items and follow-ups instead, and topic trackers can catch phrases that signal frustration, such as threats to cancel, when a team seeds a tracker with real examples.
Can PyAI analyze recordings from another phone system?
Yes, for JustCall today, with more providers listed as coming. PyAI receives finished JustCall calls through a private webhook URL, downloads each recording, transcribes it by channel or voice, and runs Recap and Trace. Duplicate pushes are removed by the provider call id before analysis starts.
PyAI is voice infrastructure. It hands you the audio and transcript, and does not replace your own models or review tools.