The Challenge
When end-customers call a Synthflow agent, every millisecond matters. A laggy transcript, a missed end-of-turn, or a mispronounced number breaks the illusion of a real conversation and breaks trust with the caller. To deliver that quality globally, Synthflow needed a speech stack that could:
- Run in real time at production scale, with hundreds of concurrent live streams and peaks of thousands of calls per minute.
- Handle conversational signals natively, including turn detection, interruptions, and end-of-turn confidence, alongside raw transcription.
- Cover the languages Synthflow customers actually sell and support in: English, German, Spanish, French, Italian, Dutch, Portuguese, and more.
- Deploy where their customers needed it, self-hosted in the EU for data residency and in the cloud for elastic burst capacity.
- Be backed by an engineering-led partner that could move at the pace of an early-stage, fast-growing platform.
Why Deepgram
After evaluating the major STT and TTS providers, Synthflow standardized on Deepgram as the speech engine behind their voice agent platform. Four reasons stand out.
1. Real-time speech-to-text built for voice agents
Deepgram’s Nova-3 model gives Synthflow industry-leading word error rates and the streaming latency a live agent needs to feel natural. Flux, Deepgram’s conversational STT, goes further, exposing turn detection, interruption handling, and end-of-turn confidence as first-class signals, so Synthflow can build agents that listen the way a person does, not the way a transcription pipeline does.
2. One model, ten languages with Flux Multilingual
With Flux Multilingual, Synthflow gets monolingual-grade accuracy across English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch through a single streaming connection. Native code-switching means a caller can move between, say, German and English mid-sentence without dropped words. For a platform serving European enterprises, that’s the difference between “it works in the demo” and “it works in production.”
3. Self-hosted deployment, EU endpoint, and Aura-2 TTS
Synthflow runs Deepgram self-hosted on Google Cloud, keeping audio inside their own infrastructure for performance, cost control, and data-residency commitments to their European customers. With Deepgram’s EU endpoint and self-hosted Flux and Aura-2 TTS, Synthflow can offer a fully EU-resident voice stack for STT, TTS, and voice agent orchestration without compromising latency.
4. A hands-on engineering partnership
Synthflow’s engineering team works directly with Deepgram’s product, model, and Applied Engineering teams in a shared Slack channel, from latency tuning and GPU sizing on L4 instances to early access on new languages and roadmap input on features that matter to voice-agent builders. With Premium Plus support and a dedicated EMEA-based account team, Synthflow gets answers in their own time zone, on their own timeline.
The Solution
A typical Synthflow voice agent flows audio in and out of Deepgram across the entire turn:
- Inbound audio from telephony or WebRTC streams into Deepgram self-hosted Nova-3 or Flux for sub-second, conversational STT.
- Flux’s end-of-turn and interruption signals drive Synthflow’s agent orchestration layer, so the LLM responds at the right moment and yields when the caller speaks over it.
- Responses are synthesized through Deepgram Aura-2 TTS, with a growing library of European voices for a natural, on-brand sound.
- All of it runs in Synthflow’s GCP environment with Deepgram cloud as elastic overflow for traffic spikes.
Results
- Tens of thousands of audio hours processed per month across the Synthflow platform, with hundreds of concurrent live streams and peaks of thousands of calls per minute.
- Sub-second turnaround on real-time STT, enabling voice agents that feel conversational in live interactions.
- EU-resident deployment unlocked for German and broader European customers via self-hosted Deepgram and the EU endpoint.
- Faster language rollout: Flux Multilingual’s ten-language launch, including German, Spanish, French, Italian, Dutch, and Portuguese, lets Synthflow expand to new markets without rebuilding the speech stack.
- A consolidated speech stack: Synthflow runs STT and TTS through a single partner, simplifying procurement, observability, and on-call.
Looking Ahead
Synthflow and Deepgram continue to push the envelope on what real-time voice agents can do, with deeper language coverage (including additional European and MENA languages), tighter end-of-turn behavior, expanded Aura-2 voices, and joint engineering on self-hosted scale. As Synthflow grows into new verticals and regions, Deepgram remains the speech engine underneath.
