Running Voice AI on Local, Self-Hosted Models: The Complete Guide
What running voice AI on your own hardware actually buys you, with measured latency, honest benchmark numbers and the stack worth running today.
Read the full guide →Deep Dives in This Series
Live vs Recorded Speech-to-Text: Which One Does a Voice Agent Need?
Live speech-to-text hears the caller during a voice agent call, and recorded speech-to-text reads the saved audio afterwards. What each one needs, Nvidia's two Nemotron models as the example, and where each runs in Dograh.
Nvidia Nemotron 3.5 Transcribe: What It Means for Voice Agents
What Nvidia's open-weight Nemotron 3.5 Transcribe does, why it is built for recordings rather than live calls, and how to run it on a voice agent's call recordings on your own servers.
STT and TTS Concurrency Limits: When Several API Keys Help
Where speech-to-text and text-to-speech vendors count concurrency, what several API keys on one Dograh connection do, when separate keys add capacity, how to turn a limit into a call count, and when to ask the vendor instead.
Voice Cloning for Pre-Recorded and TTS Playback in One Voice
How to make a voice agent's recorded lines and generated speech sound like one person: get written consent, clone the speaker with a TTS provider, put the recordings and the voice ID in Dograh, and tell callers it is AI.
How to Run Kokoro TTS Locally for a Phone Voice Agent
A step-by-step tutorial to run Kokoro TTS on your own GPU with Kokoro-FastAPI inside Dograh's Docker setup, set the model, voice and speed fields, and prove the voice with a browser call and a phone call.
How to Run Faster-Whisper Locally for a Voice Agent With Speaches
A step-by-step tutorial to run faster-whisper on your own GPU with Speaches, inside Dograh's Docker setup, connect it as the agent's speech-to-text, pick a model size for phone audio and prove it with a test call.
Kokoro vs Chatterbox vs Coqui XTTS: Open-Source TTS for Voice Agents
A comparison of Kokoro, Chatterbox and Coqui XTTS for a self-hosted voice agent: licenses, how each connects to Dograh, voice quality and speed on a phone line, and voice cloning.
Whisper vs Parakeet vs Voxtral: Choosing a Self-Hosted STT Model
A comparison of Whisper, Parakeet and Voxtral for a self-hosted phone voice agent: how each model is served, accuracy on phone audio, speed and GPU memory, languages and licenses, and how each connects to Dograh.
How to Choose an LLM for Voice Agents: Test the Tool Calls First
How to choose a language model for a voice agent: test tool calls made from spoken input first, then reply speed on a real phone call, then cost per call, and decide between hosted models and open weights, with the models Dograh can run.
BYOK vs Managed Models: Controlling Voice AI Spend
Bring your own key (BYOK) versus a platform's managed models for voice AI: who bills you, what each rate covers, how key pools and backup models control spend and outages, what BYOK does not change about data, and how to set up each mode in Dograh.
How to Cut Voice AI Costs by Mixing Pre-Recorded Audio With TTS
Record the lines your agent repeats, synthesise the rest in the same cloned voice, and see which line on the bill actually moves.
Why Voice Data Needs to Stay Local: Privacy, PII, and Biometric Risk
A phone call is ordinary personal data and a biometric identifier at the same time. What Illinois, the EDPB and the 2026 voiceprint filings mean for where your inference runs.
How to Build a Voice Agent on Fully Local Open-Source Models
The build side of running voice AI locally: what fits on one 24GB card, which runtime to use per stage, and what the latency actually looks like on a phone call.