tflite_pruned_int8

Tiny Wav2letter Pruned INT8

Description

Tiny Wav2letter is a tiny version of the original Wav2Letter model. It is a convolutional speech recognition neural network. This implementation was created by Arm, pruned to 50% sparsity, fine-tuned and quantized using the TensorFlow Model Optimization Toolkit.

License

Apache-2.0

Network Information

Network Information	Value
Framework	TensorFlow Lite
SHA-1 Hash	edc581b85190b2bcbfba904b50645264be52f516
Size (Bytes)	3997112
Provenance	https://github.com/ARM-software/ML-zoo/tree/master/models/speech_recognition/wav2letter
Paper	https://arxiv.org/abs/1609.03193

Performance

Platform	Optimized
Cortex-A	✔️
Cortex-M	✔️
Mali GPU	✖️
Ethos U	✔️

Key

• ✔️ - Will run on this platform.
• ✖️ - Will not run on this platform.

Accuracy

Dataset: Fluent Speech (trianed on LibriSpeech,Mini LibrySpeech,Fluent Speech)
Please note that Fluent Speech dataset hosted on Kaggle is a licensed dataset.

Metric	Value
LER	0.0283
WER	0.089

Optimizations

Optimization	Value
Quantization	INT8

Network Inputs

Input Node Name	Shape	Description
input_1_int8	(1, 296, 39)	Speech converted to MFCCs and quantized to INT8

Network Outputs

Output Node Name	Shape	Description
Identity_int8	(1, 1, 148, 29)	A tensor of time and class probabilities, that represents the probability of each class at each timestep. Should be passed to a decoder. For example ctc_beam_search_decoder.