mozilla-deepspeech-0.8.2

Use Case and High-Level Description

The mozilla-deepspeech-0.8.2 model is a speech recognition neural network pre-trained by Mozilla based on DeepSpeech architecture (CTC decoder with beam search and n-gram language model) with changed neural network topology.

For details on the original DeepSpeech, see paper.

For details on this model, see repository.

Specification

Metric

Value

Type

Speech recognition

GFlops per audio frame

0.0472

GFlops per second of audio

2.36

MParams

47.2

Source framework

TensorFlow*

Accuracy

Metric

Value

Parameters

WER @ LibriSpeech test-clean

8.39%

with LM, beam_width = 32, Python CTC decoder, accuracy checker

WER @ LibriSpeech test-clean

6.13%

with LM, beam_width = 500 , C++ CTC decoder, accuracy checker

WER @ LibriSpeech test-clean

6.15%

with LM, beam_width = 500 , C++ CTC decoder, demo

NB : beam_width=32 is a low value for a CTC decoder, and was used to achieve reasonable evaluation time with Python CTC decoder in Accuracy Checker. Increasing beam_width improves WER metric and slows down decoding. Speech Recognition DeepSpeech Demo has a faster C++ CTC decoder module.

Use accuracy_check [...] --model_attributes <path_to_folder_with_downloaded_model> to specify the path to additional model attributes. path_to_folder_with_downloaded_model is a path to the folder, where the current model is downloaded by Model Downloader tool.

Input

Original Model

  1. Audio MFCC coefficients, name: input_node, shape: 1, 16, 19, 26, format: B, N, T, C, where:

    • B - batch size, fixed to 1

    • N - input_lengths (see below) - number of audio frames in this section of audio

    • T - context frames: along with the current frame, the network expects 9 preceding frames and 9 succeeding frames. The absent context frames are filled with zeros.

    • C - 26 MFCC coefficients per each frame

    See <omz_dir>/models/public/mozilla-deepspeech-0.8.2/accuracy-check.yml for all audio preprocessing and feature extraction parameters.

  2. Number of audio frames, INT32 value, name: input_lengths, shape 1.

  3. LSTM in-state (c) vector, name: previous_state_c, shape: 1, 2048, format: B, C.

  4. LSTM input (h, a.k.a hidden state) vector, name: previous_state_h, shape: 1, 2048, format: B, C.

When splitting a long audio into chunks, these two last inputs must be fed with the corresponding outputs from the previous chunk. Chunk processing order must be from early to late audio positions.

Converted Model

  1. Audio MFCC coefficients, name: input_node, shape: 1, 16, 19, 26, format: B, N, T, C, where:

    • B - batch size, fixed to 1

    • N - number of audio frames in this section of audio, fixed to 16

    • T - context frames: along with the current frame, the network expects 9 preceding frames and 9 succeeding frames. The absent context frames are filled with zeros.

    • C - 26 MFCC coefficients in each frame

    See <omz_dir>/models/public/mozilla-deepspeech-0.8.2/accuracy-check.yml for all audio preprocessing and feature extraction parameters.

  2. LSTM in-state vector, name: previous_state_c:0, shape: 1, 2048, format: B, C.

  3. LSTM input vector, name: previous_state_h:0, shape: 1, 2048, format: B, C.

When splitting a long audio into chunks, these two last inputs must be fed with the corresponding outputs from the previous chunk. Chunk processing order must be from early to late audio positions.

Output

Original Model

  1. Per-frame probabilities (after softmax) for every symbol in the alphabet, name: logits, shape: 16, 1, 29, format: N, B, C, where:

    • N - number of audio frames in this section of audio

    • B - batch size, fixed to 1

    • C - alphabet size, including the CTC blank symbol

    The per-frame probabilities are to be decoded with a CTC decoder. The alphabet is: 0 = space, 1…26 = “a” to “z”, 27 = apostrophe, 28 = CTC blank symbol.

    NB : logits is probabilities after softmax, despite its name.

  2. LSTM out-state vector, name: new_state_c, shape: 1, 2048, format: B, C. See Inputs.

  3. LSTM output vector, name: new_state_h, shape: 1, 2048, format: B, C. See Inputs.

Converted Model

  1. Per-frame probabilities (after softmax) for every symbol in the alphabet, name: logits, shape: 16, 1, 29, format: N, B, C, where:

    • N - number of audio frames in this section of audio, fixed to 16

    • B - batch size, fixed to 1

    • C - alphabet size, including the CTC blank symbol

    The per-frame probabilities are to be decoded with a CTC decoder. The alphabet is: 0 = space, 1…26 = “a” to “z”, 27 = apostrophe, 28 = CTC blank symbol.

    NB : logits is probabilities after softmax, despite its name.

  2. LSTM out-state vector, name: cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/GatherNd (for new_state_c ), shape: 1, 2048 , format: B, C`. See Inputs.

  3. LSTM output vector, name: cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/GatherNd_1 (for new_state_h), shape: 1, 2048, format: B, C. See Inputs.

Download a Model and Convert it into OpenVINO™ IR Format

You can download models and if necessary convert them into OpenVINO™ IR format using the Model Downloader and other automation tools as shown in the examples below.

An example of using the Model Downloader:

omz_downloader --name <model_name>

An example of using the Model Converter:

omz_converter --name <model_name>

Demo usage

The model can be used in the following demos provided by the Open Model Zoo to show its capabilities: