Operations

Call quality

Responsiveness, reliability and voice quality of live agent calls, broken down by pipeline stage and by engine.

Sample data. 1 of the calls in this view are synthetic fixtures (dataset “sample”), not real calls.

4 calls · 41 turns · judged at p90

Time to first audio

Off target

4.24 s p90

p50 3.53 s · p95 4.75 s · max 6.08 s

End of caller speech → first agent audio frame · target ≤ 1.50 s, red > 3.00 s

End-to-end response

Off target

7.47 s p90

p50 4.47 s · p95 7.92 s · max 8.95 s

Caller starts speaking → first agent audio · target ≤ 2.50 s, red > 4.50 s

Turn reliability

Off target

14.6% failed turns

Fallback turns 0.0%
No transcript 4.9%

15 of 41 turns answered

Bottleneck

Off target

Agent / LLM

45% of response time · p50 1.36 s · p90 2.51 s

Largest mean stage between end of speech and first audio

Responsiveness trend

Time to first audio

p50p90

Turns per bucket · amber/red bars exceed the failure thresholds

Where the time goes

Stage breakdown

ASRp50 244 msp90 416 ms29%
Agent / LLMp50 1.36 sp90 2.51 s45%
TTSp50 877 msp90 991 ms25%
Transcodep50 0 msp90 16 ms0%

End-of-speech capture (VAD silence window) p50 1.18 s is counted in end-to-end only.

Reliability

Outcomes and fallbacks

responded15 · 37%
interrupted1 · 2%
superseded17 · 41%
no transcript2 · 5%
failed5 · 12%
abandoned1 · 2%

ASR fallback

0.0%

TTS / voice fallback

0.0%

Agent fallback

0.0%

Barge-ins

46.3%

Voice & intelligibility

Quality from Voice Studio evaluations

No scored evaluations in this window

Run a scenario in Voice Studio → Evaluations (requires the GPU evaluation worker) to populate MOS, WER/CER and speaker similarity.

Engine comparison

Meridian vs candidates on the same KPIs

Engine	Mode	Turns	Own stage p50	Own stage p90	Δ vs Meridian	First audio p90	Failures	Empty / no text	WER vs ours
Meridian streaming ASRMeridianbaseline	Served	40	241 ms	375 ms	—	4.24 s	12.5%	5.0%	reference
Parakeet TDT 0.6B Hebrew v4 (NeMo)Open source	Shadow	4	414 ms	613 ms	+173 ms	not served	0.0%	0.0%	14.6%
Zipformer2 66M streaming Hebrew (sherpa-onnx)Open source	Shadow	4	228 ms	303 ms	−13 ms	not served	0.0%	0.0%	11.4%
Meridian sync ASR (faster-whisper)Meridian	Served	1	26.6 s	26.6 s	+26.3 s	—	100.0%	0.0%	reference

Shadow engines transcribe the same caller audio after the fact and never affect the call. “WER vs ours” is disagreement with Meridian’s transcript (no human reference); use scripted evaluations below for absolute WER/CER.

A/B · Meridian (control)

3.82 s

first audio p90 · ASR p50 415 ms · 1 calls · failures 0.0%

A/B · Candidate

—

first audio p90 · ASR p50 — · 0 calls · failures —

Drilldown

Recent calls

Started	Agent	Turns	First audio p50	Worst	Failures	Engines	Status
20 Oct, 15:55	b0742d09-a9e1-5862-8c21-33fdbf3e9b19	19	3.81 s	6.08 s	0	Meridian streaming ASR · Default (unassigned) voice	Off target
20 Oct, 15:04	b0742d09-a9e1-5862-8c21-33fdbf3e9b19	6	—	—	1	Meridian streaming ASR · Default (unassigned) voice	Off target
20 Oct, 12:25	b0742d09-a9e1-5862-8c21-33fdbf3e9b19	11	—	—	5	Meridian streaming ASR · Default (unassigned) voice	Off target
16 Oct, 07:58sample	reception-he	5	2.74 s	3.85 s	0	Meridian streaming ASR · Chatterbox Multilingual V3	Off target

Engine swap

Routing per project or agent

Transcription routing

Serves callers

Shadow on the same audio (max 3, never affects the call)

A/B candidate (per call)Split

Using Meridian defaults · served now: Meridian streaming ASR

Voice routing

Serves callers

A/B candidate (per call)Split

Using Meridian defaults · served now: Chatterbox Multilingual V3

Candidates are used only when the media runtime has an endpoint for them (MEDIA_ASR_CANDIDATES / MEDIA_TTS_CANDIDATES). Otherwise the call stays on Meridian’s engine and the plan is reported as degraded.

Registry

Engines available for comparison

Transcription (ASR)

Meridian streaming ASRMeridianproduction
internal · Hebrew: multilingual
Meridian sync ASR (faster-whisper)Meridianfallback
internal · Hebrew: multilingual
ivrit.ai Whisper large-v3-turbo (CT2)Open sourcecandidate
Apache-2.0 · Hebrew: fine-tuned · 809M · published WER eval-d1 5.3%, eval-whatsapp 7.1%
ivrit.ai Whisper large-v3 (CT2)Open sourcecandidate
Apache-2.0 · Hebrew: fine-tuned · 1.55B · published WER eval-d1 5.1%, eval-whatsapp 7.2%
Zipformer2 66M streaming Hebrew (sherpa-onnx)Open sourcecandidate
CC-BY-4.0 · Hebrew: native · 66M · published WER eval-d1 5.3%, eval-whatsapp 8.8%
Parakeet TDT 0.6B Hebrew v4 (NeMo)Open sourcecandidate
CC-BY-4.0 · Hebrew: fine-tuned · 617M · published WER eval-d1 7.9%, eval-whatsapp 11.4%
Caspi 1.7B (Qwen3-ASR Hebrew fine-tune)Open sourceresearchnon-commercial
CC-BY-NC-4.0 · Hebrew: fine-tuned · 1.7B · published WER eval-d1 4.2%, eval-whatsapp 6.0%
Meta Omnilingual ASROpen sourcecandidate
Apache-2.0 · Hebrew: multilingual · 300M-7B

Voice (TTS)

Chatterbox Multilingual V3Meridianproduction
MIT · Hebrew: multilingual · 500M
Default (unassigned) voiceMeridianfallback
internal · Hebrew: multilingual
F5-TTS Hebrew v2Open sourcecandidate
Apache-2.0 · Hebrew: fine-tuned · 300M
Phonikud TTS (ONNX)Open sourcecandidate
CC-BY-4.0 · Hebrew: native · small
Mamre Hebrew TTSOpen sourceresearchnon-commercial
Apache-2.0 code / non-commercial weights · Hebrew: native
Qwen3-TTS 1.7BOpen sourcepreview
Apache-2.0 · Hebrew: none · 1.7B
CosyVoice 3 0.5BOpen sourcepreview
Apache-2.0 · Hebrew: none · 500M