The FlexSR advantage

---

• Building and training a large vocabulary model for conventional ASR is regarded as requiring 250+ hours of representative speech to be processed - A difficult and timely project
• The speech data for a conventional ASR model needs human transcription to train the model. Requiring expertise and between 4 to 10 times the speech time
• The combined end-to-end training time is significant for conventional ASR

FlexSR can add any words and phrases to its Lexicon described in the standard International Phonetic Alphabet (IPA) format. It requires no model training

• The logistics of obtaining enough representative speech data for training a conventional ASR, typically 250+ hours is significant
• Speech data is often subject to data security restrictions in moving and processing on cloud services
• The time to achieve this for ASR is significant

FlexSR can work on any words and phrases simply described in the IPA format and doesn't need to collect sensitive speech data

• Conventional ASR still needs model training for each new word or phrase added
• Although the data volume is smaller than the large vocabulary training, adding to conventional ASR still has similar constraints on collection
• The time to perform this training and testing can be significant

FlexSR enables any words in any language to be added on an ad-hoc basis with only the IPA description required

FlexSR can analyse the Linguistic features seen in a signal against words and phrases the speaker is attempting, to determine the speaker's pronunciation variation to the ideal, for the word that has been described in the IPA format

• Conventional ASR requires new models to be trained for each variation in pronunciation with all of the associated constraints 

FlexSR can add-in the learnt or known pronunciation variation of a speaker to its matching of the signal to words in the Lexicon, enabling greater accuracy to the result.