Speech to text with Elasticsearch
summary
Prototype turns speech into phonemes, indexes them in Elasticsearch and matches phoneme sequences against OpenOffice/CU dictionaries (English ~40k, French ~60k, German ~42k) using a shingle analyzer for phoneme bigrams and a simple statistical language model to pick likely language and candidate words. Demo uses a 50 MB offline model to emit phonemes (Sphinx/Vosk-style), shows failure when the wrong language model is loaded, and generally recovers multiword phrases but stumbles on rare tokens and homophonous cross-language matches.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:36Why talk about speech text
- 3:17Speech to text in the Mid application
- 4:29No microphone button
- 5:34Open Source
- 6:41Mozilla Deep Speech
- 7:36Demo
- 10:15How does it work
- 11:29Wrong model
- 14:17The problem
- 15:27Phones
- 18:58Cms
- 19:25Reference
- 19:40Explanation
- 21:33Index
- 25:42Statistical model
- 30:41Demonstration
related talks
15mSpeling werds egspressively with rrkurrent nuril nedwirques!
44mSome tentative first steps towards a Star Trek universal communicator
31mBuilding Large-scale Localized Language Models
29mRecognizing Speech in Python
35mTricks Learned from Scaling WhisperSpeech Models to 80k+ Hours of Speech
35mSpeech Synthesis with Tacotron2 and PyTorch
31m“Sorry, Could you repeat that again?” - Speech Recognition with Python
42mdata.bythebay.io: Shubho Sengupta,