Voice Apple Silicon Benchmarks
How long GroundSlate's voice model takes to speak on Apple Silicon, through the same synthesize_speech call an AI assistant makes: one line on its own, then six lines in a single call.
Speech speed, one line against six
Wall-clock seconds, timed through the same tool an AI assistant calls.
- One line
- One synthesize_speech call with one line of script, using a voice that already exists.
- Six lines
- One synthesize_speech call with six lines.
- The model was already loaded for both runs, so neither figure includes load time.
Apple Silicon Relative Performance Metrics
The same measurements as charts. Each chart has its own scale, and the default model is marked.
- Default
- One line
- Six lines
One line wall clock
secondsOne line realtime
timesSix lines wall clock
secondsSix lines realtime
timesRealtime above 1.00× means the Mac produced it faster than it plays.
| Model | One linerealtime | Six linesrealtime |
|---|---|---|
| 1.7B | 0.95× | 1.49× |
One line on its own ran at 0.95× realtime, a little slower than the line plays. Six lines in one call ran at 1.49×. GroundSlate speaks one request at a time, so six separate calls wait in line. For anything longer than a sentence, pass the lines together with lines=[...].
GroundSlate keeps every voice it makes. Asking for the same voice again returns the saved one in about a second.
A voice made with the 0.6B model speaks with whichever size the Mac is set to. If that is 1.7B, the audio comes from 1.7B even though the voice itself was made with 0.6B.
Makers
The companies behind these models.
Curated MLX voice models
Every model GroundSlate offers for download, including ones not measured yet.
How voice generation works on a Mac
Making a voice and speaking with it are two separate steps. When you create a voice, GroundSlate records a short reference sample from a preset or from a description you write. That sample is the voice from then on. Every line after that is spoken from your script and the sample.
This page times the speaking, because that is the part you run over and over.
The voice model size is one setting for the whole Mac. The 1.7B model uses about 11 GB of memory while it works and the 0.6B model about 2.4 GB. These figures are for 1.7B.
How these numbers were measured
- Machine
- MacBook Pro, Apple M3 Max, 64 GB unified memory
- Engine
- mlx
- Default model
- mlx-community/Qwen3-TTS-12Hz-1.7B-Base-4bit
- Models locked
- Yes
- Path
- Timed through the synthesize_speech tool, so the figure includes the queue and saving the file to the library.
- Tier
- 1.7B, set once for the whole Mac. It peaks at 11.0 GB of memory, against 2.4 GB for 0.6B.
- Voice
- Already made and saved. Making a voice is a separate one-time step and is not counted here.
Questions
- How fast is Qwen3-TTS 1.7B on a Mac?
- On a MacBook Pro, Apple M3 Max, 64 GB unified memory, one line takes 8.57s (0.95× realtime) and six lines takes 14.86s (1.49× realtime). Timed through the tool an agent calls.
- Should an agent send lines one at a time or together?
- One line on its own ran at 0.95× realtime, a little slower than the line plays. Six lines in one call ran at 1.49×. GroundSlate speaks one request at a time, so six separate calls wait in line. For anything longer than a sentence, pass the lines together with lines=[...].
- Do I pay for making a voice every time it speaks?
- GroundSlate keeps every voice it makes. Asking for the same voice again returns the saved one in about a second.
- Why does a voice made with 0.6B produce audio labelled 1.7B?
- A voice made with the 0.6B model speaks with whichever size the Mac is set to. If that is 1.7B, the audio comes from 1.7B even though the voice itself was made with 0.6B.
- Which machines are in this dataset?
- MacBook Pro, Apple M3 Max, 64 GB unified memory (Reference Mac).
- Which voice model does GroundSlate use by default?
- mlx-community/Qwen3-TTS-12Hz-1.7B-Base-4bit. Qwen3-TTS 1.7B, 2.3 GB on disk.