Training someone to change the perceived gender of their own voice is a similar process to training them to speak with a new accent. I recently prototyped a gender-affirming voice therapy app for transgender women: https://own-voice-dev.netlify.app/
So now I’m curious if our two applications use the same approach under the hood! I have a detailed tech stack rundown on this page: https://own-voice-dev.netlify.app/#/info but in brief I use OpenVoice v2 to convert someone’s voice toward a differently-gendered reference voice, but leave it at only like 30%-60% transformed so it still sounds like them and not the reference. This transformed voice then becomes a practice reference for drilling.
I know it’s bad form to discuss that, but this is a completely normative comment, with no opinion given so far from being a flame. And it adds to the technical discussion!
I think this is a super intriguing project! I’ve been experimenting with some similar work on my old unfinished songs from eons past. It’s been really fulfilling to take that old work and see the original vision spring to life through modern tools.
I think that is one of our biggest takeaways with this. The goal was never to make AI music just to get tons of streams. We were curious! And now we get to share these reimagined versions with friends and family and rekindle some great memories from back in the day.
Exactly. Even if just for Chuck and I, this was something both of us needed. It brought a lot of positivity to each of our lives for a lot of different reasons.
Hi, I'm Madison Rickert, a senior software engineer who has been shipping code to production for over 20 years. Experienced engineering leader, presently all-in on agentic engineering, shipping code daily and finding new ways to push the envelope on its capabilities.
I'm an active contributor to several open-source MCP servers, build agent-first CLI tools, solve computer vision problems, and maintain a growing library of agent skills and agent harnesses I rely on daily.
For the past five-plus years I've been a partner and the in-house technical force at a creative studio. Before that, I was a senior IC and fractional engineering lead for a variety of startups and nonprofits, working across the full stack and mentoring teams of up to 5.
I have a penchant for both business and personal finance. Across my businesses and clients, I've been responsible for managing organizational finances and countless projects with typical budgets in the 5 to 6 figure range. Over the years this led me into financial automation, and I have several open-source projects in my portfolio in areas around agentic financial automation and investment management.
I'm also a working musician, and a hyper-organized Burning Man camp lead, best known around the Northern California festival scene for my saxophone performances for crowds up to 3,000 people.
I'm pivoting back to senior IC engineering after a long stretch as partner-operator. I'm excited to keep expanding the capabilities of AI, and my creative-tech and financial background will be a real asset to the right organization. If you're building software for creative work, media systems, music tech, financial tools, or AI tooling, I'd love to hear about it.
- Strong: Using the TS/JS ecosystem to render UI and tools
- Medium: Creating custom application windows in a web view. Window management features are pretty limited currently (can’t resize or render a native “close” button on the window chrome)
- Strong: Creating custom control panels for integrating with external services (like pushing and pulling audio clips or midi to/from an external service)
- Weak: Anything real-time, that’s still the domain of Max4Live or other control APIs, so don’t expect playback automation
- Medium: Tinkering with the Ableton project, clips, tracks, etc. API surface is still incomplete. Like I can read warp markers but not create them, for example. And I can’t access the global time signature settings.
The extensions are pretty aggressively sandboxed, and I appreciate that security consciousness in this season of the js ecosystem. It’s a hassle though if you want to save or load files outside of your little sandbox folder.
Thanks a bunch for the summary! Overall seems like a great start, very similar to how Figma got started, really aggressive sandboxing, then slowly they'll open it up as they explore the exact API surfaces.
Gotta say, slightly sad about the global time signature not being available to get/set, when I read about the new SDK yesterday my mind immediately jumped to bunch of use cases which all use the global time signature :P Oh well, I'm sure it'll be available in the future.
Thank you! If I’m reading your findings correctly, this is focused on UI capabilities, and likely rides on top of the existing Live Object Model (LOM) rather than a new integration.
Which is a bummer, because LOM is fantastic for what it does, but has many many many frustrating gaps. Like, it cannot add devices to the master/main track. It can’t read MIDI pitch bend events, etc.
Hi, I'm Madison Rickert, a senior software engineer who has been shipping code to production for over 20 years. Experienced engineering leader, presently all-in on agentic engineering, shipping code daily and finding new ways to push the envelope on its capabilities.
I'm an active contributor to several open-source MCP servers, build agent-first CLI tools, solve computer vision problems, and maintain a growing library of agent skills and agent harnesses I rely on daily.
For the past five-plus years I've been a partner and the in-house technical force at a creative studio. Before that, I was a senior IC and fractional engineering lead for a variety of startups and nonprofits, working across the full stack and mentoring teams of up to 5.
I have a penchant for both business and personal finance. Across my businesses and clients, I've been responsible for managing organizational finances and countless projects with typical budgets in the 5 to 6 figure range. Over the years this led me into financial automation, and I have several open-source projects in my portfolio in areas around agentic financial automation and investment management.
I'm also a working musician, and a hyper-organized Burning Man camp lead, best known around the Northern California festival scene for my saxophone performances for crowds up to 3,000 people.
I'm pivoting back to senior IC engineering after a long stretch as partner-operator. I'm excited to keep expanding the capabilities of AI, and my creative-tech and financial background will be a real asset to the right organization. If you're building software for creative work, media systems, music tech, financial tools, or AI tooling, I'd love to hear about it.
I’ve given it access to my small business books for the last few months (attended sessions only) and so far it’s helped me clean up countless errors made by humans, at the expense of a small handful of duplicated transactions that got shaken out pretty quickly.
How do you know those duplicates are the only errors it made? You weren't aware of the apparently countless human errors before, so how would you be aware of Claude's errors?
Redundant verification is built into the workflow, both at the agent-instruction level and downstream. Reconciling a monthly bank statement with your account register in QuickBooks, for example, has the benefit of making sure you haven’t omitted transactions or added phantom ones, and once you finish reconciling a period you lock the register to prevent further changes. I supervise mutations closely and can relax my oversight when I know there are redundant checks for correctness downstream.
The class of human errors I’m encountering most often are ones caused by missing context: misfiling of transactions because my bookkeeper didn’t recognize a vendor, didn’t know a transaction needed to be attached to a specific project or client, lacked access to my calendar, didn’t have my login to pull receipts, didn’t have time to understand a spreadsheet. Claude has ready access to all of these has been remarkably adept at synthesizing that to help it come at accounting tasks with a level of detailed-oriented obsessiveness that no human I’ve ever hired has shown.
I’ll admit this article was enough to convince me to port one of my CLI tools from Python to Rust last night and I got a 30x performance gain with a binary 20% of the size. Not bad!
So now I’m curious if our two applications use the same approach under the hood! I have a detailed tech stack rundown on this page: https://own-voice-dev.netlify.app/#/info but in brief I use OpenVoice v2 to convert someone’s voice toward a differently-gendered reference voice, but leave it at only like 30%-60% transformed so it still sounds like them and not the reference. This transformed voice then becomes a practice reference for drilling.