Voice agents fail in ways a transcript cannot show. A reply lands two seconds late. An agent mispronounces a name. A caller gets cut off, and the agent continues speaking. Teams that test only the text miss all of this, and then hear about it when customers start raising tickets with your customer support.
Platforms exist to detect these failures before and after launch, such as Netra and Cekura. They both do simulated calls and scoring of production traffic. They differ in what they publish, where they go deep, and who each one fits best. Here's what I could verify, and where each platform excels.
Start with the failure you need to fix
There are three things most voice teams need. They need testing on the real callers they receive, accents and fast speakers and noise on the lines. They need to know where time was spent on a turn so that a slow response can be blamed on speech recognition, a model, a tool call, or speech synthesis. And they need the same checks running on live calls, so a release that passes testing does not quietly break in production.
This is part of what both platforms do, to varying degrees. The differences show up in the details and in the pricing.
Netra: deeper on the call, voice pricing on request
Netra's voice product is built on personas. For one, each syntax has its own voice/pace/clarity/accent/temper: For example, on the voice page, there is the frustrated caller, fast-talker (Indian English line), non-native speaker (Mexican Spanish line), rambler, confused caller, and caller on a noisy street. Each personifying dials through Netra to your agent and records the call. Most test suites have a clean script--one. This one tests the range of people who actually phone in.
There are three flavors of netra evaluators. For listening, it checks word error rate, correctness of transcription, and time to first transcript. For dialog, it measures response quality, goal completion, tool usage, and toxicity and bias. For speech, it measures pronunciation, rate of speech, language quality, and humanness. It gives each group a score for each turn.
The trace view is where Netra stands out. Each call has two sides on separate audio tracks, speech recognition, the model, tool calls, speech synthesis all timed separately as spans. If there's a slow response, you can see which stage that was. Instead of text, Netra extracts talk share, silences, overlaps, and interruptions from the recording. Other stacks can connect LiveKit or WebSocket stream in a few lines of code. Vapi agents connect from the dashboard for simulation.
Netra also features observability, testing, red teaming, and alerts for text agents. A team already running text agents can keep both channels in one place. The website lists SOC 2 Type II, HIPAA, and GDPR compliance.
There are gaps. Netra's pricing page has plans for spans, test runs, and evaluator runs. It doesn't list a voice rate, so you won't be able to estimate costs from the page with a voice team. Contact Netra support for voice testing pricing before you compare budgets. The Free plan allows you to run 25 test runs per month, sufficient to test the product.
Cekura: voice and chat, with a published voice rate
Cekura: a testing/monitoring platform for agent voice and chat It supports custom metrics, simulated calls before launch, scoring of production calls, A/B testing across agent versions, and alerts. It integrates with Retell, Vapi, ElevenLabs, LiveKit, Pipecat, and Bland. It reports percentiles, measures latency from the stereo recording, so if you have a slow tail, it will still show behind a good average.
Cekura publishes its voice meter. Third-party sites charge pay on demand at $0.25/voice-testing minute and $0.05/monitored call, up to 10 concurrent calls, plus a $500 Startup plan/month with a signed BAA. In terms of budget, a team can estimate its bill in advance.a. With its API and MCP access, it's a good choice for teams who want to run regression tests in CI. Its weaknesses are the same as Netra's. I could confirm that the one-foottrace is not so obvious at the turn level, and with high call volumes, a per-minute meter can be more expensive than a flat plan.
When Netra is the right choice
Netra is the choice when it's a question of diagnosis. Per-turn trace and the persona library are better tools for diagnosing why a given reply was late or incorrect (accents, fast speakers, noisy lines). It's also nice to have for teams already running text agents on the platform, since you get a set of traces, evaluation process, and location here for channel-convergence failures.
There are no voice minutes with Netra's Pro plan ($39/month), just spans, test runs, and experiment evaluation. Confirm the voice pricing with the Netra’s sales team to calculate the pricing.
When to Use the Cekura Advantage
Purchase Cekura if you want a predictable budget. You can predict spend with a published per-minute rate. CI integration suit teams that do frequent changes to agents and want the same set of tests on each change. As it stands, Cekura is a safer bet for those looking for a published voice rate to budget spend before a sales conversation.
The verdict
Netra is the better choice for diagnosing why calls fail. Its personas test the calls you really get, its evaluators rate the conversation/voice, and its traces trace where each turn lost. For planning voice-testing spend from a published rate, Cekura is the stronger choice today.
Pick by the failure you are trying to fix,. Then run both against the same test. Play back 20 real failed calls from your own voice agent on both and count how many of the issues your customers reported are caught. That result will tell you more than any comparison study.
Top comments (0)