Skip to content

Technology

A translation stack designed around a 500 ms budget.

Machine translation stopped being the hard part years ago. Delivering it inside the pause a human naturally leaves in conversation — that is the engineering problem, and it constrains every decision below.

End-to-end latency budget

420 ms target · 420 ms median measured

Capture & denoise30 ms
Streaming recognition120 ms
Translation180 ms
Speech synthesis90 ms

Pipeline

Every stage streams into the next.

Nothing in the chain waits for a complete sentence. Partial transcripts feed translation, partial translations feed synthesis, and the reply begins before the speaker has finished.

~30 ms

01 · Capture

A beamforming array isolates the speaker from the room. Crowd noise, music and cross-talk are stripped before a single word is transcribed.

~120 ms

02 · Understand

Streaming speech recognition transcribes partial audio as it arrives, resolving language, intent and sentence boundaries on the fly.

~180 ms

03 · Translate

A context-aware translation model carries meaning — not words — preserving tone, register, names and the jargon of your industry.

~90 ms

04 · Speak

Neural speech synthesis replies in a natural voice while the sentence is still forming, so the exchange keeps the rhythm of real conversation.

Principles

Three constraints we refuse to trade away.

Latency is the product

Anything above half a second stops feeling like a conversation and starts feeling like a phone tree. Every layer of the stack is designed backwards from that budget.

  • Streaming end-to-end — no wait for silence
  • Speculative decoding on partial transcripts
  • Edge inference in 14 regions

Context beats vocabulary

Literal translation fails exactly where it matters: contracts, diagnoses, prices, names. BabelGO carries the meaning of a sentence, grounded in the setting it was said in.

  • Domain glossaries per workspace
  • Speaker and turn-level memory
  • Register and formality preservation

Private by construction

Conversations are the most sensitive data a company holds. Ours is a pipeline you can point at an auditor without flinching.

  • Zero-retention mode by default
  • Encrypted in transit and at rest
  • EU and US data residency

Privacy

What happens to a conversation after it ends.

Nothing. That is the whole answer, and it is the default for every account.

Zero retention by default

Audio buffers are held only as long as the packet they belong to. Transcripts exist for the duration of the session and are dropped when it ends, unless a workspace explicitly opts in to storage.

Regional processing

EU traffic terminates and is processed on EU infrastructure. The same holds for US workloads. Nothing crosses a boundary that a customer has not opted into.

No training on customer speech

Customer conversations are never used to train or fine-tune models. Improvements come from licensed and synthetic corpora, and from evaluations that customers opt into explicitly.