Getting Over the Boring Stuff Quicker - Building a Semi-Automated Speech Audio Annotation Tool
summary
A semi-automated pipeline uses Mozilla DeepSpeech plus optional Google speech to text API, segments audio with voice activity detection, and runs a Flask backend with front end, persistent volume and kubernetes deployment. It auto-fills transcriptions when model confidence passes a threshold, can auto-save results, and fine-tunes the base model from collected labels then gracefully reloads new checkpoints. Demo shows faster annotation and improved accuracy after retraining but fully automatic saving remains risky for data quality.
AI-written summary · about
chapters (9)
- 0:00Start
- 0:43A Corner of My Life
- 2:24Speech Recognition
- 3:56Model Results
- 5:23Deep Learning Model Training
- 6:03Manual Annotation
- 6:23Semi-automated Annotation
- 7:14Recognition Service
- 22:38Take-away
related talks
30mProductionizing an unsupervised machine learning model to...
26mMake Python Talk, Make Python Listen
29mBuilding Machine Learning demos with Python
33mBuilding the next-generation Conversational AI with Python and Deep Learning
07mBackstage interview
13mMining Parliamentary Gold - Building Hebrew ASR
21mGPT-3 wrote the description for this talk... Scary or exciting?
49mKeynote 4: Here Come The Robots - Python and Machine Learning