In our project, DeepSpeech was used as a Speech to Text (STT) system. STT converts spoken words into text. DeepSpeech works with a neural network to transcript the spoken words into text.
- Fast install for the english version
- Multiple pretrained models with different data sources available
- Can be trained with new words really fast
- There are pretrained models available
- It does not consume much cpu load
- If you add pronouncing to english words, DeepSpeech can also understand them
- Only one german pretrained model available. This means that you have no choice between different models and can not try out which model you like best yourself
- Pretrained models are not compatible with other versions of DeepSpeech.
- For a german installation, DeepSpeech must be installed manually without help by Rhasspy
- DeepSpeech in conjunction with Rhasspy and a german language model is difficult to install. If DeepSpeech is automatically installed with Rhasspy version 2.5.7, deepspeech version 0.6.1 is installed. However, there is no german language model for this version. Therefore, we manually upgrade deepspeech to version 0.9.0. Read more about this step in this section
- As for the german model some words are sometimes not recognized. A very clear pronunciation is necessary so that the words are correctly recognized. Non-native speakers are at a distinct disadvantage here. This could also apply to english-speaking models. This has not been tested by us.
- You can not speak free to DeepSpeech, because it trains only the words you write into your sentence.ini. There is an open transcription option, but when you activate it, the transcription is really messed up.
Once DeepSpeech is set up, it works acceptably. DeepSpeech has particular problems recognising german with an accent. It can be trained well and quickly with new words. Unfortunately, for pre-trained language models other than english, there is usually little to no choice. However, it should be noted that DeepSpeech is still under development. Therefore, it can be hoped that there will be further updates in the future and that DeepSpeech will be improved in usability and functionality. It should be noted that there may be problems with speech impediments or people with a strong accent. The more unclear the person speaks, the worse DeepSpeech understands what was said.