Reinventing Speech-to-Text Transcriptions with Go and Whisper
summary
Cost-effective, open-source speech-to-text built in Go using whisper.cpp Go bindings that run a local Whisper model (trained on ~680k hours). Pipeline: download YouTube with youtube-dl, convert MP4→WAV with FFmpeg, call the tiny model via the Go bindings, produce SRT segments and store results in SurrealDB. Demo shows a CLI that transcribes short ads and caches results; it works end-to-end, but large videos slow builds (15 min → ~4 min with Docker) and demos had occasional network hiccups.
AI-written summary · about
chapters (13)
- 0:00Workshop introduction
- 4:57Go bindings overview
- 10:21Whisper and Whisper.cpp
- 15:17Speech-to-text architecture
- 18:51Setting up prerequisites and repository structure
- 21:02Basic CLI in Go
- 27:51YouTube video download
- 39:35MP4 to WAV conversion
- 47:14Docker containerization
- 58:14Transcribe with Whisper.cpp
- 1:06:22Saving transcriptions to SurrealDB
- 1:20:32Final demo
- 1:30:09Q&A
related talks
35mTricks Learned from Scaling WhisperSpeech Models to 80k+ Hours of Speech
25mPerforming Transcription of WebRTC Media Streams in Browser
26mMake Python Talk, Make Python Listen
13mMining Parliamentary Gold - Building Hebrew ASR
01h 04mAutomating Podcast Localization via OpenAI APIs
10mLightning Talk: Multiplayer Machine Learning with Metaflow, OpenAI Whisper and...
45mTranscend Language Barriers with Custom Machine Learning APIs: um...
34mText-to-Speech Data Preparation and Fine-tuning Workshop