Éanna Ó Catháin | 145c88f | 2020-11-16 14:12:11 +0000 | [diff] [blame] | 1 | # Copyright © 2020 Arm Ltd and Contributors. All rights reserved. |
| 2 | # SPDX-License-Identifier: MIT |
| 3 | |
| 4 | """Utilities for speech recognition apps.""" |
| 5 | |
| 6 | import numpy as np |
| 7 | import pyarmnn as ann |
| 8 | |
| 9 | |
| 10 | def decode(model_output: np.ndarray, labels: dict) -> str: |
| 11 | """Decodes the integer encoded results from inference into a string. |
| 12 | |
| 13 | Args: |
| 14 | model_output: Results from running inference. |
| 15 | labels: Dictionary of labels keyed on the classification index. |
| 16 | |
| 17 | Returns: |
| 18 | Decoded string. |
| 19 | """ |
Nina Drozd | 4018b21 | 2021-02-02 17:49:17 +0000 | [diff] [blame] | 20 | top1_results = [labels[np.argmax(row)] for row in model_output] |
Éanna Ó Catháin | 145c88f | 2020-11-16 14:12:11 +0000 | [diff] [blame] | 21 | return filter_characters(top1_results) |
| 22 | |
| 23 | |
| 24 | def filter_characters(results: list) -> str: |
| 25 | """Filters unwanted and duplicate characters. |
| 26 | |
| 27 | Args: |
| 28 | results: List of top 1 results from inference. |
| 29 | |
| 30 | Returns: |
| 31 | Final output string to present to user. |
| 32 | """ |
| 33 | text = "" |
| 34 | for i in range(len(results)): |
| 35 | if results[i] == "$": |
| 36 | continue |
| 37 | elif i + 1 < len(results) and results[i] == results[i + 1]: |
| 38 | continue |
| 39 | else: |
| 40 | text += results[i] |
| 41 | return text |
| 42 | |
| 43 | |
| 44 | def display_text(text: str): |
| 45 | """Presents the results on the console. |
| 46 | |
| 47 | Args: |
| 48 | text: Results of performing ASR on the input audio data. |
| 49 | """ |
| 50 | print(text, sep="", end="", flush=True) |
| 51 | |
| 52 | |
| 53 | def quantize_input(data, input_binding_info): |
| 54 | """Quantize the float input to (u)int8 ready for inputting to model.""" |
| 55 | if data.ndim != 2: |
| 56 | raise RuntimeError("Audio data must have 2 dimensions for quantization") |
| 57 | |
| 58 | quant_scale = input_binding_info[1].GetQuantizationScale() |
| 59 | quant_offset = input_binding_info[1].GetQuantizationOffset() |
| 60 | data_type = input_binding_info[1].GetDataType() |
| 61 | |
| 62 | if data_type == ann.DataType_QAsymmS8: |
| 63 | data_type = np.int8 |
| 64 | elif data_type == ann.DataType_QAsymmU8: |
| 65 | data_type = np.uint8 |
| 66 | else: |
| 67 | raise ValueError("Could not quantize data to required data type") |
| 68 | |
| 69 | d_min = np.iinfo(data_type).min |
| 70 | d_max = np.iinfo(data_type).max |
| 71 | |
| 72 | for row in range(data.shape[0]): |
| 73 | for col in range(data.shape[1]): |
| 74 | data[row, col] = (data[row, col] / quant_scale) + quant_offset |
| 75 | data[row, col] = np.clip(data[row, col], d_min, d_max) |
| 76 | data = data.astype(data_type) |
| 77 | return data |
| 78 | |
| 79 | |
| 80 | def decode_text(is_first_window, labels, output_result): |
| 81 | """ |
| 82 | Slices the text appropriately depending on the window, and decodes for wav2letter output. |
| 83 | * First run, take the left context, and inner context. |
| 84 | * Every other run, take the inner context. |
Nina Drozd | 4018b21 | 2021-02-02 17:49:17 +0000 | [diff] [blame] | 85 | Stores the current right context, and updates it for each inference. Will get used after last inference. |
Éanna Ó Catháin | 145c88f | 2020-11-16 14:12:11 +0000 | [diff] [blame] | 86 | |
| 87 | Args: |
| 88 | is_first_window: Boolean to show if it is the first window we are running inference on |
| 89 | labels: the label set |
| 90 | output_result: the output from the inference |
| 91 | text: the current text string, to be displayed at the end |
| 92 | Returns: |
| 93 | current_r_context: the current right context |
| 94 | text: the current text string, with the latest output decoded and appended |
| 95 | """ |
Nina Drozd | 4018b21 | 2021-02-02 17:49:17 +0000 | [diff] [blame] | 96 | # For wav2letter with 148 output steps: |
| 97 | # Left context is index 0-48, inner context 49-99, right context 100-147 |
| 98 | inner_context_start = 49 |
| 99 | inner_context_end = 99 |
| 100 | right_context_start = 100 |
Éanna Ó Catháin | 145c88f | 2020-11-16 14:12:11 +0000 | [diff] [blame] | 101 | |
| 102 | if is_first_window: |
| 103 | # Since it's the first inference, keep the left context, and inner context, and decode |
Nina Drozd | 4018b21 | 2021-02-02 17:49:17 +0000 | [diff] [blame] | 104 | text = decode(output_result[0][0][0][0:inner_context_end], labels) |
Éanna Ó Catháin | 145c88f | 2020-11-16 14:12:11 +0000 | [diff] [blame] | 105 | else: |
| 106 | # Only decode the inner context |
Nina Drozd | 4018b21 | 2021-02-02 17:49:17 +0000 | [diff] [blame] | 107 | text = decode(output_result[0][0][0][inner_context_start:inner_context_end], labels) |
Éanna Ó Catháin | 145c88f | 2020-11-16 14:12:11 +0000 | [diff] [blame] | 108 | |
| 109 | # Store the right context, we will need it after the last inference |
Nina Drozd | 4018b21 | 2021-02-02 17:49:17 +0000 | [diff] [blame] | 110 | current_r_context = decode(output_result[0][0][0][right_context_start:], labels) |
Éanna Ó Catháin | 145c88f | 2020-11-16 14:12:11 +0000 | [diff] [blame] | 111 | return current_r_context, text |
| 112 | |
| 113 | |
| 114 | def prepare_input_tensors(audio_data, input_binding_info, mfcc_preprocessor): |
| 115 | """ |
| 116 | Takes a block of audio data, extracts the MFCC features, quantizes the array, and uses ArmNN to create the |
| 117 | input tensors. |
| 118 | |
| 119 | Args: |
| 120 | audio_data: The audio data to process |
| 121 | mfcc_instance: the mfcc class instance |
| 122 | input_binding_info: the model input binding info |
| 123 | mfcc_preprocessor: the mfcc preprocessor instance |
| 124 | Returns: |
| 125 | input_tensors: the prepared input tensors, ready to be consumed by the ArmNN NetworkExecutor |
| 126 | """ |
| 127 | |
| 128 | data_type = input_binding_info[1].GetDataType() |
| 129 | input_tensor = mfcc_preprocessor.extract_features(audio_data) |
| 130 | if data_type != ann.DataType_Float32: |
| 131 | input_tensor = quantize_input(input_tensor, input_binding_info) |
| 132 | input_tensors = ann.make_input_tensors([input_binding_info], [input_tensor]) |
| 133 | return input_tensors |