Pre-trained DanSpeech Models¶

All of the available DanSpeech models are shown below. If you need to finetune or train your own model, then you can find more information at DanSpeech training repository.

Recommended usage for all models (except a custom model):

from danspeech.pretrained_models import TestModel
model = TestModel()

Available models¶

danspeech.pretrained_models.DanSpeechPrimary(cache_dir=None)¶

Deepest and best performing DanSpeech model.

3 Conv layers

9 RNN Layers with 1200 hidden units

Parameters: cache_dir (str) – If you wish to use custom directory to stash/cache your models. This is generally not recommended, and if left out, the DanSpeech models will be stored in the ~/.danspeech/models/ folder.
Returns: Pretrained DeepSpeech (Best Performing) model.
Return type: danspeech.deepspeech.model.DeepSpeech

danspeech.pretrained_models.TestModel(cache_dir=None)¶

Test model that runs very fast even on CPUs

Performance is very bad!

2 Conv layers

5 RNN Layers with 400 hidden units

Parameters: cache_dir (str) – If you wish to use custom directory to stash/cache your models. This is generally not recommended, and if left out, the DanSpeech models will be stored in the ~/.danspeech/models/ folder.
Returns: Pretrained DeepSpeech (Testing purposes) model
Return type: danspeech.deepspeech.model.DeepSpeech

danspeech.pretrained_models.Baseline(cache_dir=None)¶

Baseline DanSpeech model.

2 Conv layers

5 RNN Layers with 800 hidden units

Parameters: cache_dir (str) – If you wish to use custom directory to stash/cache your models. This is generally not recommended, and if left out, the DanSpeech models will be stored in the ~/.danspeech/models/ folder.
Returns: Pretrained DeepSpeech (Baseline) model.
Return type: danspeech.deepspeech.model.DeepSpeech

danspeech.pretrained_models.CPUStreamingRNN(cache_dir=None)¶

DanSpeech model with lookahead, which works as a real-time streaming model.

This model runs on most modern CPUs.

2 conv layers

5 RNN layers (not bidirectional) with 800 units each

Lookahead context is 20

Parameters: cache_dir (str) – If you wish to use custom directory to stash/cache your models. This is generally not recommended, and if left out, the DanSpeech models will be stored in the ~/.danspeech/models/ folder.
Returns: Pretrained DeepSpeech (Streaming for CPU) model
Return type: danspeech.deepspeech.model.DeepSpeech

danspeech.pretrained_models.GPUStreamingRNN(cache_dir=None)¶

DanSpeech model with lookahead, which works as a real-time streaming model.

This model will not be able to follow a stream of data on regular CPUS. Hence, use a GPU.

2 conv layers

5 RNN layers (not bidirectional) with 2000 units each

Lookahead context is 20

Parameters: cache_dir (str) – If you wish to use custom directory to stash/cache your models. This is generally not recommended, and if left out, the DanSpeech models will be stored in the ~/.danspeech/models/ folder.
Returns: Pretrained DeepSpeech (Streaming for GPU) model
Return type: danspeech.deepspeech.model.DeepSpeech

danspeech.pretrained_models.Folketinget(cache_dir=None)¶

The deepest and best performing DanSpeech model finetuned to data from Folketinget.

3 Conv layers

9 RNN Layers with 1200 hidden units

Parameters: cache_dir (str) – If you wish to use custom directory to stash/cache your models. This is generally not recommended, and if left out, the DanSpeech models will be stored in the ~/.danspeech/models/ folder.
Returns: Pretrained DeepSpeech (Folketinget tuned) model.
Return type: danspeech.deepspeech.model.DeepSpeech

danspeech.pretrained_models.TransferLearned(cache_dir=None)¶

The Librispeech English model adapted to Danish while keeping the conv layers and the lowest/first RNN layer frozen

This model performs better than the DanSpeechPrimary model on noisy data.

2 Conv layers

5 RNN Layers with 800 hidden units

param str cache_dir: If you wish to use custom directory to stash/cache your models. This is generally not recommended, and if left out, the DanSpeech models will be stored in the ~/.danspeech/models/ folder.

Returns: Pretrained DeepSpeech (Transfer learned from English) model
Return type: danspeech.deepspeech.model.DeepSpeech

danspeech.pretrained_models.EnglishLibrispeech(cache_dir=None)¶

English trained model on the Librispeech corpus.

2 Conv layers

5 RNN Layers with 800 hidden units

Parameters: cache_dir (str) – If you wish to use custom directory to stash/cache your models. This is generally not recommended, and if left out, the DanSpeech models will be stored in the ~/.danspeech/models/ folder.
Returns: Pretrained DeepSpeech (English speech recognition) model.
Return type: danspeech.deepspeech.model.DeepSpeech

danspeech.pretrained_models.CustomModel(model_path)¶

Instantiates customly trained models

Parameters: model_path (str) – Path to custom trained model
Returns: Custom DeepSpeech model.
Return type: danspeech.deepspeech.model.DeepSpeech