Pinch (YC W25)

Real-time AI models to make conversations across languages feel fluid, expressive, and human

[Christian Safka @christiansafka]

@StartPinch Dubbing now has an editor.

Preview how the model identifies speakers in your video or audio, review transcripts and translations, and make edits before generating the final dub.

[Pinch (YC W25) @StartPinch]

In 2026, real-time speech translation is no longer a feature - it's a baseline expectation.

The question is no longer whether real-time voice translation exists.

The real question: Can we carry a conversation across languages with the same ease as we would in our own?

[Pinch (YC W25) @StartPinch]

Real-time voice translation used to be:

Speech → Text → Translate → Text → Speech 🔁

But the field is shifting further.

Now it’s going end-to-end.

We are not completely at the finish line yet. The best systems today are hitting ~3-4 seconds of latency.

[Pinch (YC W25) @StartPinch]

Great models don't come from one brilliant moment.

They come from continuous iteration and learning from real-world feedback.🔁

We shipped, learned, and shipped again. And again. 🚀

Stay close folks, because something totally new is coming 👀

[Pinch (YC W25) @StartPinch]

A model can speak 100 languages and still not understand a single culture.

Multilingual - it understands the words.
Multicultural - it understands what those words carry.

Culture lives in how you say it.
At @StartPinch, we preserve the character of the speaker.

[Pinch (YC W25) @StartPinch]

7,000 languages exist. Only 5% are online.
95% of human languages are invisible to the digital world.

@UNESCO calls this a human rights issue. We agree.

Language shouldn't determine who gets to participate.
That's why we build what we build. https://t.co/mONvO7oLYz

[Pinch (YC W25) @StartPinch]

Real-time speech-to-speech translation is a problem that's not fully solved yet.

At Pinch, that's the problem we work on every day. Real conversation isn't just words - it's warmth, hesitation, emphasis, humor.

[Pinch (YC W25) @StartPinch]

The idea of universal translation is older than computers by centuries 🌍

A thread on how we got from pure philosophy to real-time speech-to-speech 🧵

[Pinch (YC W25) @StartPinch]

Researchers started asking:
What if you could skip the text step entirely?

2022: Translatotron 2 - end-to-end speech-to-speech translation. Your voice, preserved.
Then Meta, ElevenLabs, ByteDance, Kyutai all entered the race.

[Pinch (YC W25) @StartPinch]

@ComfyUI integration repo: https://t.co/jiNNOf1XAj

[Pinch (YC W25) @StartPinch]

Pinch integrations come with examples🔌

@n8n_io — example workflow as a JSON file, downloadable directly from our GitHub repository. Import it into your n8n platform!

@ComfyUI — the Pinch translation node plugs right into your ComfyUI setup with full custom configuration.

[Pinch (YC W25) @StartPinch]

@n8n_io integration repo: https://t.co/yg2oyf2qbI

[Pinch (YC W25) @StartPinch]

Github link: https://t.co/9p95U2SrRP

[Pinch (YC W25) @StartPinch]

Real-time language translation should be available, accessible, and easy to use.
We're bringing real-time language translation to @LiveKit Agents with our open-source plugin.

Get started with one line:
pip install livekit-plugins-pinch

[Pinch (YC W25) @StartPinch]

The video above is Spanish dubbed to English:

Original speech volume: 80%
Translation lag: 1s
Reduce accent: off

[Pinch (YC W25) @StartPinch]

One thing we love about Pinch Dubbing - the control it gives you!
You're not just uploading a file and getting a translation back. You can fine-tune the entire experience.
🧵👇