Can you run it?

Pick your machine. We say what works well, what is slow and what will not fit, and how sure we are.

Only the Apple M4 16 GB figures were timed. Everything else is marked estimated or untested, and we would rather say so than guess at a number.

Your machine

Apple Silicon Mac

Apple Silicon Mac, 8 GBM1 to M4 base models
Apple Silicon Mac, 16 GBWhere every number was measured (Apple M4)
Apple Silicon Mac, 24 GB or moreRoom for every model at once

Windows / Linux with an NVIDIA GPU

NVIDIA GPU, 4 GB VRAM (for example RTX 3050 Ti)Laptop GPUs such as the RTX 3050 Ti
NVIDIA GPU, 6 to 8 GB VRAMRTX 3060, 4060, 3070 class
NVIDIA GPU, 12 GB VRAM or moreRTX 3080 Ti, 4070 Ti and up

Any laptop

Any laptop, no GPU (CPU only)Intel or AMD without a dedicated GPU

With a free Groq key (cloud speech + thinking)Groq listens and thinks online. Voices still run on your machine.

Runs well

Apple Silicon Mac, 16 GB, local models only

The whole studio runs offline with no keys. A spoken reply takes about 3 seconds, and the heavy voices are for renders.

On your machine

Works well

Live voice agent
Cloning in your voice
Your voice in Indian languages
Offline thinking (no internet)
Hands-free conversation

Slow or needs extra setup

22 Indian-language preset voices

Won't fit

Nothing: it all fits.

MeasuredMeasured on an Apple M4 with 16 GB, local models only unless the line says Groq.

See every featureNotes per feature, speed, disk space, all machines side by side, and a plain-language glossary

Details for Apple Silicon Mac, 16 GB, local models only.

What works

WorksLive voice agentfull spoken reply 3.1 s
WorksCloning in your voicevoice built in 2.4 s
WorksYour voice in Indian languages5 to 7 s per sentence on the Apple GPU, measured; renders, not live calls
Works with limits22 Indian-language preset voices35 s per sentence, and it swaps with the rest
WorksOffline thinking (no internet)Llama 3.2 3B on the Apple GPU
WorksHands-free conversationfirst audio in about 1 s

Switch off, or use the cloud for

Run ./auralis warm before a demo so the first reply is about 3 s instead of about 18 s.
Use Indic-Speak and DhVaani (your voice in Indian languages) for Studio renders. Keep live calls on Kokoro voices.
Avoid loading Indic-Speak while a live call runs: with every model loaded memory hits 10 to 12 GB and the Mac swaps.

Phone calls (through Exotel, tested on a trial account) and meeting booking with Google Calendar (tested live) use the same live agent, so they work wherever "Live voice agent" works. Phone calls also need an internet address that Exotel can reach, and they sound best with a preset voice because cloned voices pause between sentences. WhatsApp voice notes are built but not yet tested live.

Expected speed and memory

Full spoken reply (median)

3.1 sMeasured

English 2.1 s, Hindi 4.1 s; 30 spoken questions, local models only

First audio (median)

1.0 sMeasured

Voice clone built

2.4 sMeasured

warm, consent check included

Indian-language preset voice

about 35 s per sentenceMeasured

Indic-Speak is about 7x slower than real time: for Studio renders, not live calls

Your voice in Indian languages

about 5 to 7 s per sentenceMeasured

DhVaani on the Apple GPU, 1.2 to 1.5x real time on a quiet M4 and up to 3x under load: for Studio renders, not live calls. Hindi alone can fall back to Chatterbox (30 to 40 s per sentence on the CPU)

Memory with every model loaded

10 to 12 GBMeasured

16 GB works but swaps under heavy use

Disk space

Core install on disk

about 15 GBMeasured

10.6 GB of models (embedding model 4.3 GB, Qwen3-TTS 2.3 GB, local language model 1.7 GB, Whisper 2.0 GB, Kokoro 0.3 GB) plus 4.1 GB of Python environments.

Add Indic-Speak (--indic)

about 7.5 GBMeasured

6.6 GB model plus a 0.9 GB environment, read on the M4; the same files on any machine. Needs about 8 GB of free memory to run.

Add Chatterbox (--chatterbox)

about 4.3 GBMeasured

3.0 GB model plus a 1.3 GB environment, read on the M4; the same files on any machine.

Add DhVaani (--dhvaani)

about 1.3 GBMeasured

0.5 GB model and vocoder plus a 0.8 GB environment, read on the M4; the same files on any machine. Needs an HF_TOKEN whose account has been granted access by ARTPARK-IISc.

All machines side by side

Showing local models only. Your pick is outlined.

Which capabilities each machine supports, with local models only.
Machine	Live agent	Cloning	Indian-language clone	22 Indian voices	Offline thinking	Hands-free
Mac 8 GBEstimatedTier: Basic	Works with limits	Works with limits	Works with limits	Not practical	Works with limits	Not practical
Mac 16 GBMeasuredTier: Runs well	Works	Works	Works	Works with limits	Works	Works
Mac 24 GB+EstimatedTier: Full studio	Works	Works	Works	Works	Works	Works
NVIDIA 4 GBUntestedTier: Basic	Works with limits	Works with limits	Works with limits	Not practical	Works with limits	Not practical
NVIDIA 6–8 GBUntestedTier: Runs well	Works with limits	Works	Works	Works with limits	Works	Works with limits
NVIDIA 12 GB+UntestedTier: Full studio	Works	Works	Works	Works	Works	Works
CPU onlyEstimatedTier: Basic	Not practical	Works with limits	Works with limits	Not practical	Not practical	Not practical

works well
works with limits (slow, swaps memory, or needs extra setup)
not practical on this machine

Measuredtimed or sized on a real machine by this project
Estimatedworked out from model sizes or research notes, not run on this hardware
Untestedno one has run auralis on this hardware yet

Words used here, in plain language

Groq: A free online service that listens to the caller and thinks of the answer, so your computer only has to speak it. Needs internet and a free key.
Local models: The AI runs on your own computer, with no internet. Nothing leaves the machine, but it needs more memory.
Kokoro: The small ready-made voice that talks during live calls and chats. Fast, and runs even without a graphics card.
Indic-Speak: Ready-made voices in 22 Indian languages. Big and slow, so it is used to prepare audio in advance, not for live calls.
DhVaani: Clones your own voice in Indian languages, from a sample you agreed to give. Made by ARTPARK-IISc. A few seconds per sentence, so it is for finished audio, not live calls.
Render: Making a finished audio file you can play later, instead of talking live.
Swapping: When a computer runs out of memory and borrows disk space. It still works, but everything gets slow.
VRAM: The memory on a graphics card. AI voices and models have to fit in it to run fast.
Offline thinking: The agent can work out answers with no internet by using an AI model stored on your machine.
Measured, estimated, untested: Measured means we timed it on a real machine. Estimated means we worked it out from model sizes. Untested means nobody has run it there yet.

Minimum: any 8 GB laptop with a free Groq key. Recommended: Apple Silicon 16 GB or an NVIDIA GPU with 6 GB+.

Setup steps, the optimisation research and the trial report with the measured numbers are in the project's docs, shared on request.