Enhance speech recognition accuracy using custom language models
summary
Amazon Transcribe’s custom language models train on domain text (emails, manuals, Trustpilot) uploaded to S3 and produce measurable gains — Octopus Energy reports WER improvements of about 12–21% using even <5 MB of data. A pipeline running Airflow on Kubernetes pulls ~1M minutes/month of calls every 15 minutes, sends audio to Transcribe, and surfaces results in Kraken via a “Kraken Karaoke” UI; CLM currently supports batch transcription with plans to move toward real time.
AI-written summary · about
chapters (27)
- 0:00Introduction
- 0:29Agenda
- 0:55Audio content
- 1:25Transcription
- 1:36Other ASR services
- 1:57Amazon Transcribe
- 3:15Transcribe features
- 5:58Demo
- 7:42Training data vs tuning data
- 9:16Guidelines
- 10:06Training the model
- 10:36Naming the model
- 11:38Training your model
- 12:02Selecting a new model
- 12:57Comparing transcription results
- 13:28Word error rate analysis
- 16:13Octopus Energy
- 18:09Kraken
- 18:55How we use Amazon Transcribe
- 19:52How we train our custom model
- 20:19Why we wanted a custom model
- 22:12Evaluation of the model
- 22:56Customer service demo
- 24:48Text analytics demo
- 25:46Where to go from here
- 26:19Datadriven business
- 26:46Next steps
related talks
33mMaximize lifetime value of media content
43mBreaking the language barrier: how do we quickly add multilanguage support in our AI application?
35mTricks Learned from Scaling WhisperSpeech Models to 80k+ Hours of Speech
01h 34mReinventing Speech-to-Text Transcriptions with Go and Whisper
45mTranscend Language Barriers with Custom Machine Learning APIs: um...
16mThe End of Awkward AI Transcriptions
45mA Pipeline to Transcribe Speech in Millions of Videos per Day
18mUsing Machine Learning to Enrich and Analyze Real-time Communications