Dubbie is an open-source AI dubbing studio that costs $0.1/min, which is about 10-30x less than alternatives like ElevenLabs, RaskAI, or Speechify. While still in early development and not at feature parity with these alternatives, Dubbie offers enough features to create dubs for basic videos.
Here we focus on the technical aspects of Dubbie. For more on motivations and mission, see why we made Dubbie.
For questions/bugs/contributions, join our Discord server.
- NextJS 14: Client app (app.dubbie.com)
- Node: Longer running functions (initialization/exporting)
- Openrouter: LLM selection for best-fit tasks
- Azure/OpenAI: Voice generation
- Prisma: Database interface (Postgres)
- Firebase: Storage
- Clerk: User authentication
- Stripe: Payments
- ShadcnUI: Some UI components
Note: These choices are more due to personal preference/experience rather than the "best".
This project is a monorepo with 4 packages
- /next
- /node
- /shared
- /db
next and node are applications that are deployed to vercel/railway.
db contains our prisma schema + client.
shared contains individual functions that are used for inside of both next and node.
You be wondering, next is frontend and node is backend, how can they use the same functions? To be honest, the code is not perfectly organized, and not all code in shared can be used in both web and node runtime. However, since NextJS is actually just a node server that's serving web pages, most of the code in shared can be used in the "Server Actions" side of NextJS. This may be a little confusing if you've not familiar with Next14, RSC and Server Actions. But it really does make things a lot easier!
- The user uploads the video and click “create project”
- Upload the video to firebase storage
- Extract the audio and upload it to firebase storage as well
- Transcribe the audio via Whisper
- Which will give us the entire transcription in a big paragraph, as well as time stamps for each word.
- Use an LLM to break down the entire paragraph into individual sentences.
- Match the individual sentences with the word level timestamps to figure out when each sentence begins and ends.
- Since the LLM output may not be “perfect” match, we will then use an approximation algorithm.
- Use an LLM to translate each sentence it into the language the user selected.
- We do this translation chunk by chunk, and use certain techniques to ensure the output matches the input.
- Use a text to speech API(currently just Azure and OpenAI) to generate audio!
- Upload those audio files to firebase storage, and save the URLs to our database via Prisma.
- The frontend client updates and renders all of that so users can preview realtime and edit
On a high level: there are 3 elements that we need to sync
- Video element
- Timeline scrubber
- Invisible audio player
Tone.js connects individual audio URLs and serves as the main timer. See useAudioTrack.ts for implementation details.
- client bugs
- Regenerating segment and moving segments sometimes crashes the Next.js app. Unclear why. But there's some sorts of race condition happening.
- backend scale
- currently there's only one instance of the backend. There are no limits/queues either. So it's almost guarantee'd that if many people uses it at once, it'll crash. I don't have too much experience with this, so if you wanna help that'd be very much appreciated :)