CEDX Speech · AI
CEDX Speech transcribes batch and live audio, scores every job for word error rate and confidence, and runs voice agents whose interruptions and CSAT are watched as closely as their costs.
The demo's field-radio channel runs at an estimated 17.8% word error rate with 0.61 confidence — and it sits on the overview as a priced alert, not buried in a log. Hard audio is the case this product is built for.
Runs on demo data. The budget card says “illustrative from the demo sample” on the screen itself.
What it is
320 jobs in view with model, status, word error rate, confidence, duration and cost on each row. job_18713 — Claims intake on listen-accurate — came out at 6.8% WER and 0.83 confidence; the Support IVR job beside it ran 14.0% at 0.65. The difficult rows are not hidden: a 22.4% WER job sits flagged Warning in the same table.
48 streaming sessions with an average first byte of 527ms and an average RTF of 1.03 — wall time against audio time, printed on the header. A 13-minute field-radio stream holds 0.71 confidence; the errors, one at 0.33, are in the table with everything else.
14 voice agents, 23,734 sessions in 30 days, and the columns that make voice honest: handle time, barge-in share, tools per session and CSAT. booking.confirmation carries 6,980 sessions at 4.8 CSAT despite a 32% interrupt rate; dispatch.status-line is paused at 3.8 — which is also a decision this table supports.
Product tour
Not a mockup and not a concept deck. This is what opens at /app/speech.
01 — Overview
320 jobs in the sample — 256 completed, 39 failed, 14 processing, 11 queued — with the 12.2% fail rate on a headline card. Storm week's field-radio spike is annotated on the chart, and the alerts are priced: accurate-model spend up 22% week over week at $39.40, field-radio WER risk at $22.10.
02 — Transcripts
Each job row names its project and model with status, WER, confidence, duration and cost. The most expensive visible job — 18:58 of Support IVR on listen-accurate — cost $0.4552; the field-radio warnings cost $0.0000 and say why they're flagged anyway.
03 — Live
Forty-eight sessions with RTF, first-byte, duration, confidence and cost per row. The fastest first byte in view is 261ms; the average is 527ms. Sessions that errored keep their rows — an 0:58 stream at 0.50 confidence is right there.
04 — Voice agents
Every agent pairs a listen model with a speak voice — listen-balanced with speak-warm, listen-stream with speak-deep — and the table watches sessions, handle time, barge-in, tools per session and CSAT. ivr.support-router averages 5:10 a call; utility.meter-report interrupts only 4% of the time.
Who runs it
Roles, not references. We have no named customers yet, so nobody in these photographs is quoted, credited or claimed as one.
Watches the Support IVR project: 18-minute calls at 14% WER, the es-US confidence gate at 0.88, and the disconnect cluster the alert traces to packet loss.
Support IVR · 0.88 gate
Reads transcripts against the audio: sorts by WER, opens the 22.4% rows, and decides what the glossary has to learn next.
median WER · 10.9%
Keeps the retention question answerable: every job has an id, a model, a duration and a cost, exportable to CSV.
320 jobs · exportableThe shape of it
Every figure below is legible in the captures above. Nothing here is a projection of your estate — it is the state of the demo data.
How it runs
Audio arrives as a batch job or a live stream; the demo mixes both — 55 hours across 320 jobs in the sample window.
The model choice is a stated trade: listen-fast is cheap, and listen-accurate crossed a budget alert on Aug 2 — the chart annotation says when.
WER, confidence and gates turn quality into a queue: field radio's 17.8% estimate, the es-US IVR's 0.81 against its 0.88 gate.
Voice agents close the loop — a listen model and a speak voice per agent, with handle time, barge-in and CSAT watched per session.
One record
Speech is not a bolt-on phone bot. Its projects in the demo — Support IVR, Claims intake, Dispatch notes — are named for the work the rest of the estate does.
Listen and speak models are metered per job and per session here; per-call cost and rate discipline is what the gateway keeps for the whole estate.
CEDX DeskThe demo's Support IVR project transcribes exactly what a desk team hears — 18-minute calls and all.
CEDX CopilotAsk AI sits on every Speech screen too, and the alerts it would summarize — WER risk, budget crossings — are the ones priced on this page.
Limits
Finding this out on the third call is worse for you than reading it here, and worse for us.
Start
Open it right now
Your audio, your gates
Speech with the rest of it
Questions
Yes. Every screenshot is a capture of the running build and you can open the same build at /app/speech. It runs on demo data — Harbor Field Services is the demo workspace, not a customer.
Word error rate — the share of words a transcript gets wrong, lower is better. The demo's median on completed jobs is 10.9%; its best flagged cell runs 5.1% on clean phone audio, and its worst alert is 17.8% on a field-radio channel. The honest answer is that it depends on the audio, which is why the number is printed per job.
Both. The Live screen shows 48 streaming sessions with an average first byte of 527ms and an average RTF of 1.03 — wall time against audio time — and the Transcripts screen carries the batch jobs. The overview counts both in the same 55 hours.
They stay visible. 39 of the 320 sample jobs failed — 12.2%, on a headline card — and the failed rows keep their model, duration and cost in the table. The Live screen keeps its 8 errored sessions for the same reason.
No certification has been issued, and no call-recording or regulated-industry compliance is claimed here. What we can evidence about hosting, encryption, tenant isolation and production access is written up on the security page.
Live build, demo data, no card. Then ask what your worst audio channel would score.