AI Researcher, Voice Models

About Freya

Voice AI for Enterprises

Tech

About the Role You will own the backbone that runs thousands of concurrent AI-powered phone calls inside bank-grade on-premise environments. Not just keeping pods alive, you architect distributed systems that handle real-time voice, scale STT / LLM / TTS inference across customer GPU clusters, integrate with enterprise telephony (Cisco CUBE, Genesys, Asterisk), and deploy behind the firewalls of largest financial institutions. Your work decides whether our platform answers a bank's rush-hour traffic or leaves customers on dead air. What You'll Do • Own on-prem deployments into OpenShift clusters inside banks. Helm charts, image registries, GPU allocation, CyberArk integration, SAML 2.0 / OIDC SSO. • Scale GPU inference infrastructure for our STT, TTS, and LLM models across multiple customer environments (H100 / H200, NVLink, Triton or vLLM). • Integrate with telephony: Asterisk, SIP trunks, Cisco CUBE, Genesys, WebRTC. SIP header parsing (X-Genesys-*), direction routing, warm transfers, DTMF. • Own reliability: Splunk SIEM forwarding, Langfuse and Grafana observability, incident playbooks for bank-grade 24/7 SLAs. • Security and compliance: RBAC, pentest remediation, KVKK and BDDK compliance patterns, pod security policies. • Scale with growth: we onboard a new bank or insurer every quarter. Each is a new on-prem environment with its own constraints. • Spot flaws early. We are building new architecture for a regulated industry. You help us see what needs to be solved next. Interesting Problems to Own • On-prem meets streaming. Most voice AI stacks assume cloud. We run the same stack inside banks with zero internet egress. Novel problems in image delivery, model updates, secrets rotation. • Bank-scale concurrency. A single campaign can put millions of customers on the line the same afternoon. Queueing, graceful degradation, GPU-aware autoscaling are yours to design. • Legacy-meets-new telephony. Cisco CUBE, Asterisk, Genesys, SIP, WebRTC. You wrangle old-school protocols alongside modern streaming stacks. What Makes You a Great Fit • 3+ years building and scaling distributed systems. Deep Kubernetes / OpenShift knowledge. AWS / GCP helpful. • Fundamentals plus. You can sketch how a SIP INVITE flows through a proxy, explain a K8s GPU scheduler, or tell us the obscure thing you fell asleep reading last night. • Real-time systems experience. Low-latency streaming or inference. Voice / video is a big plus. • On-prem mentality. You have shipped software into environments you did not fully control: Turkish bank, European healthcare, US regulated finance, anything similar. • Startup hats. You have worked where problems find you before process does. • Opinionated without alienating. Opinions drive progress, but you find compromises with customers and teammates. • Familiar with: Kubernetes / OpenShift, Helm, Docker, Terraform, NVIDIA GPU stack, Asterisk, SIP, Cisco CUBE, Genesys, Splunk, Grafana, CyberArk, Python or Go. Bonus Points • Telephony systems (SIP, VoIP, WebRTC). • ML infrastructure, model serving, or GPU computing (Triton, vLLM, TensorRT). • Real-time audio processing. • Banking / fintech / BDDK and KVKK compliance familiarity. • Fluent Turkish or comfortable working with Turkish customer teams daily. How to Apply Email dev@freyavoice.ai with a short note about the hardest infra problem you have shipped and your CV. We move fast, expect a reply within 48 hours. Don't worry about the checklist. We hire for how you think, not how many boxes you tick. The work is hard, the hours are long, and most of what we ship nobody has built before. If that sounds good instead of scary, apply.

The role

We build the most realistic voice models in the world. Our English model, Freya TTS Adam V1, ranks #1 among AI models on Design Arena's Audio Realism Bench, behind only real human recordings: https://www.designarena.ai/leaderboard/audiorealismbenchWe own the full voice stack: our own speech-to-text and text-to-speech models, LLM serving, and a real-time agent runtime. Our models run in production, often on-prem inside banks and insurers, and answer real customer calls every day.You will build voice models that are the best in the world for the languages and conditions we serve. That means 8 kHz telephone audio, noise, accents, people talking over each other, and answers that must come back in a few hundred milliseconds.This is not a paper-only role. You take an idea to a training run, then to a model that serves live calls. You measure it on real traffic.## What you will do- Train and fine-tune speech-to-text models that hold up on telephone audio, noise, accents and code switching.- Build text-to-speech models that sound human: natural prosody, correct question intonation, stable voices across long calls.- Work on audio codecs and speech-language models that listen and speak in one system.- Collect, clean and filter large speech datasets, and design the training curriculum.- Run ablations that isolate one change. Report objective metrics and listening tests side by side.- Make models fast. Quantization, streaming inference, latency on a single GPU.## Who we are looking for- You have trained STT, TTS, codec or speech-language models yourself, not only used them.- You can read a new paper on Monday and have a working experiment by Wednesday.- You design experiments that answer one question and drop ideas when the data says no.- You care about how a model sounds on a real call, not only about a benchmark number.- You own the work from research to deployment.## Bonus points- Papers, open source models or popular repos in speech or language AI.- Experience with distributed training on many GPUs.- Experience with real-time audio, telephony or streaming systems.- Experience with Turkish or other low-resource languages.## What we offer- Real data and real users. Your model hears customer calls the week it ships.- Dedicated GPUs for training and serving.- A small team where you decide what to build next.- Competitive salary and meaningful equity.If you want to build the best voice models in the world and see them work on live calls, we want to hear from you.

Skills

• Torch/PyTorch
• Machine Learning
• Deep Learning
• Speech Recognition