I forked a version that works on the Jetson Orin Nano Super...

[illiastra]

Firstly, this is my first contribution to Github and I hope I'm not stepping on anyone's toes or doing anything wrong as far as attribution is concerned. My assumption is that in forking from travisvn's codebase, the attribution is embedded by that action.

Anyway...

Claude and I created a version of this project that runs in the highly-constrained memory of a Jetson Orin Nano Super. While the throughput it quite low, it does work. It's not anything you'd want to use for an AI assistant due to the latency, it would be perfectly fine for transcribing long-form texts into audio or pre-baking transcriptions for later use.

The startup sequences moves a memory around quite a bit but, once started, it is quite reliable.

Aside from working on the Orin Nano Super, there should be no noticeable difference.

[illiastra]

Here's a log that got generated through the use of the web-based front end, with the text field being:

This is a test of Chatterbox TTS running on the Jetson Orin Nano Super.

It took about 20 seconds for it to generate the audio.

===============================================================================

🎵 Using default voice for OpenAI voice 'alloy' (no alias mapping)
📊 Request #6 - Initial memory: CPU 3285.3MB, GPU 1861.8MB allocated
Processing 1 text chunks with configured sample and parameters:

• Exaggeration: 0.5
• CFG Weight: 0.5
• Temperature: 0.8
  Generating audio for chunk 1/1: 'This is a test of Chatterbox TTS running on the Je...'
  WARNING:root:Reference mel length is not equal to 2 * reference token length.

✓ Audio generation completed. Size: 210,318 bytes
🧹 Memory cleanup (collected 1940 objects)
📊 Request #6 - Final memory: CPU 3287.6MB, GPU 1862.3MB allocated
📈 Memory change: CPU +2.4MB, GPU +0.4MB
INFO: 172.18.0.3:54922 - "POST /v1/audio/speech/upload HTTP/1.0" 200 OK