Can a DGX Spark do speech to text locally?

Yes. Qwen3-ASR recognizes speech in 30 languages on the box's CPU, behind an OpenAI-compatible audio route, and answers the official OpenAI SDK as it is.

Does speech slow down the model that serves?

No. It uses five small CPU cores for each engine and no GPU memory. With both speech engines busy, the flash lane's decode did not move (47.5 tok/s).

Which audio files does it read?

Any audio ffmpeg reads, such as a phone's m4a or a browser's webm. Long recordings are cut into pieces of 30 s at their quiet moments.

Is speech part of the one command?

Not yet. It is opt-in: from a clone, after ./install.sh, run ./install-speech.sh. It takes about 3 minutes and is proved before it ends. The one command's updates leave it as it is, so run ./install-speech.sh again after each one.

Speech recognition and text to speech on a DGX Spark, on the CPU

Run ./install-speech.sh after the install: Qwen3-ASR recognizes speech in 30 languages and Kokoro-82M speaks with 41 voices in seven languages, behind OpenAI-compatible audio routes. It runs on the CPU alone, so it sits beside whichever lane serves.

How do I install speech on my Spark?

./install-speech.sh    # from a clone, after ./install.sh: about 3 minutes, proved before it ends

Speech is opt-in and not in the one command yet. It installs two engines behind OpenAI-compatible audio routes on 127.0.0.1:30024, on the box.

What does it recognize, and what does it say?

Qwen3-ASR recognizes speech in 30 languages.
Kokoro-82M speaks with 41 voices in seven languages.

Both answer the official OpenAI SDK as it is.

Does it take memory from the models?

A little. It runs on the CPU alone, five small cores for each engine, and allocates no GPU memory. But on a GB10 the CPU and the GPU share one memory: running, the lane holds 2.5 GB at rest and 2.7 GB at work with the 0.6B recognizer (4.2 GB with the 1.7B), memory that a flash lane's long prompts also want. That is why it does not start at boot unless SPEECH_BOOT=1. With both speech engines busy, the flash lane's decode did not move (47.5 tok/s).

Which recordings can I send?

Any audio ffmpeg reads, a phone's m4a or a browser's webm among them. Long recordings are cut into pieces of 30 s at their quiet moments.

Where are the details?

The languages, every number and the limits are in docs/speech-lane.md.

Speech recognition and text to speech on a DGX Spark

How do I install speech on my Spark?

What does it recognize, and what does it say?

Does it take memory from the models?

Which recordings can I send?

Where are the details?

Install it on your Spark

Questions