One session, not N adapters
Voice, tokens, and webhooks share a session model and one SDK.
Voice and AI products
Stop rewriting the codec layer for every provider
HZRelay sits between your product and the providers that carry audio, tokens, and webhooks. Twilio’s mulaw, Deepgram’s PCM, and a speech engine’s audio become one session. It is in private preview.
If you are building a product that talks, the integration work is not the model. It is the plumbing between whoever carries the call and whoever transcribes, thinks, and speaks. Twilio sends mulaw at 8 kHz. Deepgram wants PCM at 16 kHz. ElevenLabs returns PCM. A browser client wants something else again. Each webhook provider has its own JSON. Teams spend the month on codecs and reconnects.
HZRelay is that routing layer. Voice, token streams, and webhooks normalize to one session model and one SDK. You keep the product logic. It is in private preview, not a self-serve signup.
The three problems it is for
Voice paths that do not share a format. The relay transcodes and reconnects so your application code is not a pile of one-off adapters. Token streams from an LLM that have to reach more than one subscriber without you reimplementing fan-out and backpressure. Webhooks, from Stripe or GitHub or the next provider, that you do not want to parse N times M.
The inbound side we talk about in the product includes Twilio, Plivo, WebRTC, raw WebSocket, and webhooks such as Stripe and GitHub. The outbound side includes ElevenLabs and Cartesia for speech, Deepgram for transcription, and OpenAI and Anthropic for the model, plus your own app. Those are providers the routing layer is built to sit between. Using the logo is not a partnership announcement.
What you still own
The prompt, the tool calls, the decision to interrupt, and what the product is allowed to do on someone’s behalf. HZRelay does not pick those. A session exposes events you would otherwise scrape out of three vendor dashboards: speech start and end, partial and final transcripts, first token, first audio, a barge-in that flushes the speech buffer, and an adapter error that says which provider failed.
Latency is visible as stages on a session rather than as a single “it felt slow.” The product page walks a path from audio received, through speech-to-text, the model, and speech synthesis, to audio sent back. Use that when you are arguing about which vendor to swap, not as a promise that every call will land on the sample numbers.
What this product is not
It is not a call center for your clinic and it is not InterviewRoom. InterviewRoom is a finished interview a candidate can take. HZRelay is for the team building its own real-time product. FlowTux will not route mulaw. If the pain is internal tickets in Slack, you are on the wrong page.
Private preview means you talk to us before you depend on it. There is no published monthly price on the product page. Ask, and we will say what the preview includes. Do not plan a launch date on an assumption that preview means generally available.
Providers the routing layer sits between
These marks name products people in this work already use. They are not Mecverse customers and they are not endorsements.
Twilio
PSTN audio in, often mulaw at 8 kHz
Deepgram
Transcription that wants PCM, not the carrier’s format
ElevenLabs
Speech audio back toward the caller
OpenAI
Token streams that need to fan out
Anthropic
The other model path on the product page
GitHub
Webhooks that should not each get a custom parser
What changes
Built around the work, not the buzzwords
Teams building a real-time voice or streaming product
You keep the product logic
Prompts, tools, and what the agent is allowed to do stay in your code.
Private preview
There is no self-serve price on the product page. Talk to us before you date a launch on it.
The solution
The pieces that make the workflow work
Transcode on the voice path
mulaw 8 kHz in, PCM 16 kHz where the transcriber wants it, PCM back from the speech engine.
Events you can log
Speech start and end, partial and final transcripts, first token, first audio, barge-in, adapter errors.
Webhooks without a new parser each time
Stripe and GitHub shaped events are part of the inbound story on the product page.
Not a helpdesk
Internal Slack tickets are FlowTux. A finished interview is InterviewRoom.
How we put it to work
01
Name the providers
Who carries the call, who transcribes, who speaks, who thinks.
02
Put one path through the relay
A single session you can log beats a diagram of six vendors.
03
Swap a vendor without rewriting the app
That is the reason to be on a routing layer. Preview means we do this with you, not via a signup form.
Other solutions
Recruitment & staffing
Run the first interview before a recruiter gets on a call
Education & training
Close the register without rebuilding it in a spreadsheet
Software & IT teams
One queue for the Slack message, the Sentry error, and the GitHub issue
Customer support
Answer the questions you already answer every day
Offices & distributed teams
Attendance, the broken laptop, and the request that died in chat
Campus hiring
One first round for the whole batch
Clinics and practices
Answer the reception questions. Refuse the clinical ones.
Computer maintenance
A maintenance contract that says what is included
Questions
Before you get started
Is HZRelay generally available?
No. It is in private preview. Ask what the preview includes before you plan a date.
Does it replace ElevenLabs, Deepgram, or OpenAI?
No. It routes between them and your app. You still pay those providers.
Can a clinic use this to answer patients?
No. A clinic wants the WhatsApp assistant on the clinics page. HZRelay is for a team building its own real-time product.
Let’s find the right first step
Tell us what your team is trying to improve. We’ll recommend a focused setup instead of selling you a stack you do not need.
Book a conversation