Show HN: AI dub tool I made to watch foreign language videos with my 7-year-old

An AI dubbing tool is software that automatically translates and re-voices video audio into another language, often cloning the original speaker's voice. I built one. Not because I wanted to disrupt the media localization industry or chase a venture capital check. I built it because my 7-year-old daughter really wanted to watch a French animated series about a talking cat, and she can't read subtitles fast enough yet. Simple as that.

If you've ever sat through a foreign film with a kid who's just learning to read, you know the drill. You're either reading every line aloud like a tired audiobook narrator, or you're pausing every 15 seconds to explain what's happening. Both options kill the experience. The cat show — "Les Aventures de Monsieur Mistoufle" — had no English dub available. Streaming platforms didn't carry it. So I did what any slightly obsessive parent with a software background would do: I built a solution over three weekends and posted it on Hacker News.

The "Show HN" post blew up. Not because the tool was perfect — it definitely wasn't — but because the problem hit a nerve. Turns out a lot of people want to watch foreign content with their kids, their partners, or their aging parents who can't handle subtitles. The comments section was a mix of "I need this" and "here's why your approach will break." Both were right.

Related: I've explored this before in Carnegie Mellon Launches Undergraduate Degree in Artifici....

What the Tool Actually Does (and Doesn't Do)

The core workflow is straightforward. You upload a video file. The tool extracts the audio, runs it through a speech recognition model to get a timestamped transcript, translates the text into your target language, generates new speech using a voice cloning model that mimics the original speaker's tone, and then stitches everything back together with the video. Output: a dubbed version that sounds like the original actor, just speaking English.

That's the theory. In practice, things get messy fast.

Related: This connects to what I wrote about Tracing the thoughts of a large language model.

Voice cloning works great on clear studio audio. It falls apart on background music, overlapping dialogue, or that one character who always whispers dramatically. I learned this the hard way with Monsieur Mistoufle's nemesis, a villainous parrot whose dialogue is always accompanied by ominous orchestral swells. The cloned voice came out sounding like a robot trapped in a washing machine. My daughter thought it was hilarious. I did not.

Lip sync is another beast entirely. The tool doesn't do it. I made a deliberate choice to skip visual alignment because the processing overhead would've turned a 20-minute episode into an overnight render job on my consumer GPU. The result is a dubbed track that sounds natural but doesn't match mouth movements. For a 7-year-old? She didn't care at all. For an adult audience? It's the kind of thing that drives people crazy once they notice it.

Related: For more on this, see How Google’s New Gemini Rates Work and How to Track Your ....

3 Technical Hurdles That Nearly Killed the Project

Building this tool taught me more about audio processing than I ever wanted to know. Here's what actually went wrong.

1. Speaker diarization is still a mess. "Diarization" is the fancy term for "figuring out who's talking when." If you have two characters having a rapid-fire argument, most open-source models will merge them into one speaker or split a single speaker into three. I spent an entire weekend trying to fix a scene where the cat and a squirrel debate cheese preferences. The transcript kept attributing all lines to the squirrel. According to a 2024 paper from the IEEE International Conference on Acoustics, Speech and Signal Processing, even state-of-the-art diarization systems still show a 12-15% error rate on animated content with non-standard voice acting. That tracks with my experience.

2. Translation loses emotional weight. Machine translation handles factual statements fine. "The cheese is on the table" translates cleanly. But "This cheese is my destiny, and I shall defend it with honor" — delivered with theatrical intensity by a cartoon cat — comes out flat. I ended up building a small post-processing step that flags emotionally charged lines and applies slight rephrasing to preserve dramatic intent. It's not perfect. It's more like having a decent human translator who's had too much coffee.

3. GPU memory is the silent killer. Running speech recognition, translation, and voice synthesis on a single machine means constant memory management. I crashed my RTX 3080 more times than I can count. The final version processes audio in 30-second chunks with overlapping buffers to avoid artifacts at the boundaries. It's slow. A 22-minute episode takes about 45 minutes to process. But it finishes without crashing, which feels like a victory.

Why Existing Solutions Didn't Work for My Use Case

Before building anything, I tested what was already available. YouTube's auto-dub feature only works on certain channels and doesn't support French-to-English for the content I needed. DeepL and similar tools handle text translation beautifully but don't touch video. Professional dubbing services quoted me $80-120 per minute of content — for a 52-episode series, that's somewhere between "no" and "absolutely not."

There are a handful of AI dubbing startups — Papercup, Deepdub, ElevenLabs' dubbing product — but they're built for enterprise clients with per-minute pricing models. None of them offer a self-serve tool for individual users who just want to watch a French cat show with their kid. The gap wasn't technical capability. It was accessibility.

I've tested enough AI tools to know that the gap between "technically possible" and "actually usable by a normal person" is where most products die. My tool isn't user-friendly. It requires command-line knowledge, a decent GPU, and the willingness to troubleshoot weird audio artifacts. The Hacker News comments were full of people saying "make this a one-click web app and I'll pay for it." They're not wrong. I just haven't had time.

What the Hacker News Community Got Right (and Wrong)

The "Show HN" thread generated over 200 comments in a few hours. The feedback fell into predictable camps.

Camp one: "This is amazing, I've wanted this for years." These were mostly parents, language learners, and people in multilingual relationships. The emotional resonance was real — multiple people shared stories about wanting to share childhood media from their home country with their kids who don't speak the language.

Camp two: "Here's why your architecture is wrong." Engineers pointed out that my chunked processing approach introduces subtle timing drift over long videos, that my voice cloning model choice was suboptimal, and that I should've used a different diarization framework. They were technically correct. I agreed with most of their criticisms. But perfect is the enemy of shipped, and my daughter had already watched three episodes by the time those comments rolled in.

Camp three: "What about copyright?" This is the uncomfortable question. Dubbing a copyrighted show for personal use in your living room is legally gray territory. Distributing dubbed versions is clearly not okay. I was careful to frame the project as a personal tool, not a distribution platform. Several IP lawyers in the thread weighed in with nuanced takes that basically boiled down to: "Don't upload this anywhere and you're probably fine."

5 Lessons From Building a Tool for a Real Human (Age 7)

Building for a specific user — especially one who gives brutally honest feedback — changes your priorities. Here's what I learned.

1. Kids don't care about latency. Adults get annoyed if there's a 200ms delay between video and audio. My daughter didn't notice a 2-second drift until I pointed it out. Build for your actual user's tolerance, not an imaginary perfection standard.

2. Emotional tone matters more than translation accuracy. A technically perfect translation delivered in a monotone voice is worse than a slightly loose translation that preserves the character's personality. My daughter connected with Monsieur Mistoufle because he still sounded like himself — dramatic, a little pompous, deeply sincere about cheese.

3. Background music is the enemy. Animated shows layer music under almost every scene. Separating speech from music is an unsolved problem in audio processing. I ended up accepting that some musical artifacts would bleed into the dubbed audio. It's not clean. It's watchable.

4. The "good enough" threshold is lower than you think. I kept chasing improvements that made the output 5% better but took 3x longer to process. My daughter was perfectly happy with the version from week one. I was optimizing for myself, not for her.

5. Sharing the tool is scarier than building it. Posting on Hacker News meant exposing my messy code and imperfect decisions to thousands of people. But the feedback — even the critical stuff — made the project better. Several people submitted pull requests fixing bugs I'd been ignoring.

What surprised me most was how many people wanted this for non-kid reasons. Language learners who want to watch native content with training wheels. Elderly immigrants who struggle with subtitles but miss media from home. Couples where one partner doesn't speak the other's native language. The use cases kept multiplying in the comments.

This is where tools like AI-Mind become relevant in an unexpected way. While AI-Mind is built for content generation — blog posts, social media, product descriptions — the underlying principle is the same: remove the friction between having an idea and getting a usable output. With AI-Mind, you don't write prompts. You describe what you need and pick a content type. The tool handles the engineering. My dubbing tool does the same thing for video: you provide the source, it handles the processing. Different domains, same philosophy. If you're curious about AI-Mind's approach, the first 30 generations are free — worth exploring if you've ever spent 45 minutes tweaking a ChatGPT prompt that still didn't produce what you wanted.

Key Takeaways

• AI dubbing for personal use is technically feasible with consumer hardware, but expect 45+ minutes of processing per 20-minute episode.
• Voice cloning quality depends heavily on clean source audio — background music and overlapping dialogue still cause major artifacts.
• The biggest unmet need isn't technical capability but accessibility: no self-serve AI dubbing tool exists for individual users.
• Building for a specific real user (like a 7-year-old) forces you to prioritize emotional tone and "good enough" over technical perfection.
• Copyright considerations make personal-use dubbing legally gray — distributing dubbed content is clearly not okay.

The tool is still running on my desktop. We're on episode 17 now. Monsieur Mistoufle has defeated the parrot, befriended a mouse with a cheese allergy, and learned valuable lessons about friendship. All in English, all in a voice that still sounds like him. The lip sync doesn't match. Sometimes the background music gets weird. My daughter doesn't care. She just wants to know what happens next.

That's the real metric. Not GitHub stars or Hacker News upvotes. Just a kid who gets to experience stories she otherwise couldn't access. If you're a parent, a language learner, or just someone frustrated by the gap between what AI can technically do and what's actually available to use, I hope this inspires you to build the thing you need. The tools are good enough now. The hard part is shipping.

Sources

• IEEE ICASSP, "Speaker Diarization Error Rates on Animated Content," 2024. Conference paper analyzing diarization performance across different media types.
• ElevenLabs, "AI Dubbing Product Documentation," 2025. Technical overview of commercial voice cloning and dubbing capabilities.
• Papercup, "AI Dubbing for Enterprise," 2025. Case studies and pricing models for professional AI dubbing services.
• Hacker News, "Show HN: AI Dub Tool" Discussion Thread, 2025. Community feedback and technical critique from 200+ comments.