Together AI says Wonderful’s voice models reach first sentence in 274ms
A new customer case study details EU-only hosting, production-concurrency tests and a shift from engineer-assisted changes to self-service operations.
Loading page…
A new customer case study details EU-only hosting, production-concurrency tests and a shift from engineer-assisted changes to self-service operations.
Listen to this story
Together AI’s customer case study details how Wonderful moved its custom multilingual voice inference onto dedicated EU-hosted GPUs while retaining direct control over endpoint capacity, configuration and model releases. Wonderful reports 274ms to first sentence at production concurrency, near its 300ms target, and says self-service deployment has eliminated infrastructure-change tickets; the figures are customer-reported and vendor-published, and the zero-failure result applies to hundreds of calls per test run, not a stated uptime period.
EU data residency is mandatory for Wonderful’s voice endpoints; Together provides regional GPU infrastructure while Wonderful manages endpoint settings.
The latency work combines GPU clusters near Wonderful’s servers with a custom TensorRT engine and quantization.
Wonderful defines time to first sentence as when speech can begin naturally, not the time to complete a caller exchange.
Wonderful’s custom voice models can reach the point where speech begins in 274 milliseconds while handling concurrent production workloads, according to a customer case study Together AI published. The account describes more than a speed result: it explains how Wonderful runs those models entirely in EU data centers and manages deployment changes without waiting for Together’s engineers.
Wonderful builds and fine-tunes multilingual voice models for languages it says general-purpose speech models do not adequately serve. That creates two linked jobs: improving language coverage and keeping the resulting models fast enough for a live conversation. The deployment uses Together AI’s Dedicated Model Inference service, with dedicated NVIDIA GPUs running the models.
Location is a requirement, not simply a performance choice. Wonderful says every endpoint serving this voice workload must remain in the EU to meet its data-residency rules. Together supplies regional GPU infrastructure, while Wonderful controls the creation, size, scaling policies and configuration of those endpoints—the access points its software uses to reach the models.
The speed work spans placement and software. The GPU clusters sit near Wonderful’s own servers, and the optimization work includes a custom TensorRT engine and quantization. Together also says it matched Wonderful’s production configuration for tool calling: the behavior that lets a voice agent hand work to connected systems during a conversation.
Wonderful’s reported result at production concurrency, against its stated 300ms target.
The case study defines time to first sentence as the point where a text-to-speech engine can begin speaking naturally. Wonderful’s reported result sits near its 300ms target. The measure therefore concerns the start of spoken output, rather than the duration of a complete exchange with a caller.
Wonderful also reports zero failed requests across hundreds of calls per test run at production concurrency levels. That is a test-run result, not a stated duration of uninterrupted service. Both the latency and failure figures are customer results published by the infrastructure vendor; they should be read with that attribution attached.
The operational change addresses a different kind of delay. Previously, configuration changes and capacity requests involved manual edits by a Together engineer, a submitted request and a wait for turnaround. Wonderful wanted to make frequent configuration changes and run experiments without routing each infrastructure adjustment through that process.
The self-service setup puts several tasks directly in Wonderful’s hands:
The dashboard addresses a gap Wonderful had specifically flagged: seeing infrastructure problems and failed inference requests quickly enough to investigate endpoint health and take recovery action itself. That visibility matters alongside deployment access. The team wanted to inspect a problem directly, rather than escalate through support simply to learn what was happening.
Wonderful says it has filed zero tickets for scaling, model rollouts or testing since moving those tasks to self-service. Its stated payoff is more engineering time for fine-tuning voice models in the languages customers need, rather than managing the GPUs beneath them.
Loading discussion...
Join the conversation
Explain what else, if anything, would earn your trust.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.