What is DanSpeech?¶
DanSpeech is an open-source Danish speech recognition (speech-to-text) python package based on the PyTorch deep learning framework. It was developed as part of a Master’s thesis at DTU compute by Martin Carsten Nielsen and Rasmus Arpe Fogh Jensen, supervised by Professor Lars Kai Hansen.
All DanSpeech models are end-to-end DeepSpeech 2 models, trained on danish data with a CTC loss. The models are trained with various data agumentations to multiply the rather small amount of public speech recognition data available in Danish.
The models may be combined with a language model through beam-search decoding to achieve the best results, DanSpeech provides language models trained on a large danish corpus as part of the released package.
While DanSpeech models perform state-of-the-art speech recognition in Danish, performance is not perfect, and results are conditioned on specific use-cases.
Danspeech provides:
An easy-to-use Recognizer that supports different use-cases for Danish speech recognition.
Pre-trained models of varying sizes and complexities.
Pre-trained language models.
Motivation¶
We believe that speech recognition in Danish should be freely available for everyone to use. Therefore we decided to develop an open-source, and easy-to-use automatic speech recognition system for Danish.
We believe that an open-source solution can play an important role in ensuring that Danish speech recognition systems are not continually out-shined by English systems.
Speech recognition will inevitably be a big part of future IT innovations. And without an easy-to-use, and free system, innovative spirits with a desire to utilize speech recognition in the development of Danish technologies might be hindered by cost barriers.
As such DanSpeech can be used commercially by companies without the resources to develop their own speech recognition systems, or companies who simply do not wish to outsource this part of their pipeline. Deploying DanSpeech models instead of using an external API will, in addition to reducing costs, also reduce latency drastically, if deployed locally with a GPU.
Performance¶
We benchmarked the system on two Danish benchmarks, namely the publicly available Nordisk Språkteknologi (NST) dataset and our own DanSpeech dataset (~1000 noisy recordings). The performance is evaulated in Word Error Rate (WER).
Dataset |
Models |
Performance |
|---|---|---|
NST test |
DanSpeechPrimary + DSL5Gram (not pruned) LM |
12.85% WER |
DanSpeech dataset |
TransferLearned + DSL5Gram (not pruned) LM |
25.75% WER |
DanSpeech Demo¶
To test the DanSpeech models on your own data (both pre-recorded and streaming audio is supported), we have created a demo that runs as a development django server on localhost. It is easy to install and hence easy to test the models with a GUI (little technical knowledge is required to play around with the demo).
The demo also features a demo of a DanSpeech model adopted/finetuned to transcribe meetings from Folketinget (The Danish Parliament), which demonstrates the power of finetuning models to specific domains.
For more info of the DanSpeech demo, see Demo.
Train or Finetune DanSpeech models¶
If you require better performance than what is apparent from the DanSpeech pre-trained models, we’ve also created a github repository (danspeech_training), where you can train completely new models from scratch or finetune existing DanSpeech models to your specific domain/use-case (recommended method).
Training new, or finetuning, DanSpeech models is useful if you have specific knowledge about the domain you wish to apply Danish speech recognition to, and you have either domain specific text resources or, in the best case, speech data available.
Finetuning a DanSpeech model can result in much better performance but does require a certain level of technical expertise and a GPU for training. As an example of performance for such a system, see Demo.
For more information, see DanSpeech training repository.