sonusai 0.15.8__py3-none-any.whl → 0.16.0__py3-none-any.whl

sonusai/utils/asl_p56.py

@@ -22,7 +22,7 @@ def asl_p56(audio: AudioT) -> float:
     # Hangover time in seconds
     H = 0.2
     # Rounded up to next integer
-    I = np.ceil(H * SAMPLE_RATE)
+    H_samples = np.ceil(H * SAMPLE_RATE)
 
     # Margin in dB, difference between threshold and active speech level
     M = 15.9
@@ -40,7 +40,7 @@ def asl_p56(audio: AudioT) -> float:
     a = np.full(thresh_num, -1)
 
     # Hangover counter for each threshold
-    h = np.full(thresh_num, I)
+    h = np.full(thresh_num, H_samples)
 
     # Long-term level square energy of audio
     sq = sum(np.square(audio))
@@ -55,7 +55,7 @@ def asl_p56(audio: AudioT) -> float:
             if q[k] >= c[j]:
                 a[j] = a[j] + 1
                 h[j] = 0
-            elif h[j] < I:
+            elif h[j] < H_samples:
                 a[j] = a[j] + 1
                 h[j] = h[j] + 1
             else:

sonusai/utils/asr.py

@@ -1,10 +1,22 @@
 from dataclasses import dataclass
 from typing import Any
+from typing import Callable
 from typing import Optional
 
 from sonusai.mixture import AudioT
 
 
+@dataclass(frozen=True)
+class ASRData:
+    audio: AudioT
+    whisper_model: Optional[Any] = None
+    whisper_model_name: Optional[str] = None
+    device: Optional[str] = None
+    cpu_threads: Optional[int] = None
+    compute_type: Optional[str] = None
+    beam_size: Optional[int] = None
+
+
 @dataclass(frozen=True)
 class ASRResult:
     text: str
@@ -16,8 +28,24 @@ class ASRResult:
     asr_cpu_time: Optional[float] = None
 
 
+def get_available_engines() -> dict[str, Callable[[ASRData], ASRResult]]:
+    from importlib import import_module
+    from pkgutil import iter_modules
+
+    module = import_module('sonusai.utils.asr_functions')
+    engines = {method: getattr(module, method) for method in dir(module) if not method.startswith('_')}
+    for _, name, _ in iter_modules():
+        if name.startswith('sonusai_asr_'):
+            module = import_module(f'{name}.asr_functions')
+            for method in dir(module):
+                if not method.startswith('_'):
+                    engines[method] = getattr(module, method)
+
+    return engines
+
+
 def calc_asr(audio: AudioT | str,
-             engine: Optional[str] = 'deepgram',
+             engine: Optional[str] = 'aaware_whisper',
              whisper_model: Optional[Any] = None,
              whisper_model_name: Optional[str] = 'tiny',
              device: Optional[str] = 'cpu',
@@ -43,15 +71,14 @@ def calc_asr(audio: AudioT | str,
 
     from sonusai import SonusAIError
     from sonusai.mixture import read_audio
-    from sonusai.utils import asr_functions
-    from sonusai.utils.asr_functions.data import Data
+
+    available_engines = get_available_engines()
+    if engine not in available_engines:
+        raise SonusAIError(f'Unsupported ASR function: {engine}')
 
     if not isinstance(audio, np.ndarray):
         audio = copy(read_audio(audio))
 
-    data = Data(audio, whisper_model, whisper_model_name, device, cpu_threads, compute_type, beam_size)
+    data = ASRData(audio, whisper_model, whisper_model_name, device, cpu_threads, compute_type, beam_size)
 
-    try:
-        return getattr(asr_functions, engine)(data)
-    except AttributeError:
-        raise SonusAIError(f'Unsupported ASR function: {engine}')
+    return available_engines[engine](data)

sonusai/utils/asr_functions/__init__.py

@@ -1,6 +1 @@
 from .aaware_whisper import aaware_whisper
-from .aixplain_whisper import aixplain_whisper
-from .deepgram import deepgram
-from .fastwhisper import fastwhisper
-from .google import google
-from .whisper import whisper

sonusai/utils/asr_functions/aaware_whisper.py

@@ -1,8 +1,8 @@
+from sonusai.utils import ASRData
 from sonusai.utils import ASRResult
-from sonusai.utils.asr_functions.data import Data
 
 
-def aaware_whisper(data: Data) -> ASRResult:
+def aaware_whisper(data: ASRData) -> ASRResult:
     import tempfile
     from math import exp
     from os import getenv

sonusai/utils/asr_manifest_functions/__init__.py

@@ -4,3 +4,4 @@ from .librispeech import collect_librispeech_transcripts
 from .librispeech import get_librispeech_manifest_entry
 from .vctk_noisy_speech import collect_vctk_noisy_speech_transcripts
 from .vctk_noisy_speech import get_vctk_noisy_speech_manifest_entry
+from .mcgill_speech import get_mcgill_speech_manifest_entry

sonusai/utils/asr_manifest_functions/mcgill_speech.py

@@ -0,0 +1,29 @@
+from sonusai.utils.asr_manifest_functions import PathInfo
+
+
+def get_mcgill_speech_manifest_entry(entry: PathInfo, transcript_data: list[str]) -> dict:
+    from os.path import splitext
+    from os.path import basename
+    from subprocess import check_output
+
+    from sonusai import SonusAIError
+
+    name = splitext(entry.abs_path)[0]
+    duration = float(check_output(f'soxi -D {entry.abs_path}', shell=True))
+    # i.e., from MA01_02.wav, get 01_02
+    promptname = basename(name)[2:]
+    # paragraph num
+    pnum = int(promptname[0:2])
+    snum = int(promptname[3:5])
+    idx = 11 * (pnum - 1) + (snum - 1)
+    try:
+        # remove prompt-id prefix and \n suffix
+        text = transcript_data[idx][6:-1]
+    except IndexError:
+        raise SonusAIError(f'Could not find {promptname}, idx {idx} in transcript data')
+
+    return {
+        'audio_filepath': entry.audio_filepath,
+        'text':           text,
+        'duration':       duration,
+    }

sonusai/utils/audio_devices.py

@@ -0,0 +1,41 @@
+import pyaudio
+
+
+def get_input_device_index_by_name(p: pyaudio.PyAudio, name: str = None) -> int:
+    info = p.get_host_api_info_by_index(0)
+    device_count = info.get('deviceCount')
+    for i in range(0, device_count):
+        device_info = p.get_device_info_by_host_api_device_index(0, i)
+        if name is None:
+            device_name = None
+        else:
+            device_name = device_info.get('name')
+        if name == device_name and device_info.get('maxInputChannels') > 0:
+            return i
+
+    raise ValueError(f'Could not find {name}')
+
+
+def get_input_devices(p: pyaudio.PyAudio) -> list[str]:
+    names = []
+    info = p.get_host_api_info_by_index(0)
+    device_count = info.get('deviceCount')
+    for i in range(0, device_count):
+        device_info = p.get_device_info_by_host_api_device_index(0, i)
+        device_name = device_info.get('name')
+        if device_info.get('maxInputChannels') > 0:
+            names.append(device_name)
+
+    return names
+
+
+def get_default_input_device(p: pyaudio.PyAudio) -> str:
+    info = p.get_host_api_info_by_index(0)
+    device_count = info.get('deviceCount')
+    for i in range(0, device_count):
+        device_info = p.get_device_info_by_host_api_device_index(0, i)
+        device_name = device_info.get('name')
+        if device_info.get('maxInputChannels') > 0:
+            return device_name
+
+    raise ValueError('No input audio devices found')

sonusai/utils/calculate_input_shape.py

@@ -13,13 +13,12 @@ def calculate_input_shape(feature: str,
     """
     from pyaaware import FeatureGenerator
 
-    # num_classes is irrelevant, set to 2
-    fg = FeatureGenerator(feature_mode=feature, num_classes=2)
+    fg = FeatureGenerator(feature_mode=feature)
 
     if flatten:
-        in_shape = [fg.stride * fg.num_bands]
+        in_shape = [fg.stride * fg.feature_parameters]
     else:
-        in_shape = [fg.stride, fg.num_bands]
+        in_shape = [fg.stride, fg.feature_parameters]
 
     if timesteps > 0:
         in_shape.insert(0, timesteps)

sonusai/utils/create_timestamp.py

@@ -0,0 +1,5 @@
+def create_timestamp() -> str:
+    """Create a timestamp."""
+    from datetime import datetime
+
+    return datetime.now().strftime('%Y%m%d-%H%M%S')

sonusai/utils/{trim_docstring.py → docstring.py}

@@ -28,3 +28,23 @@ def trim_docstring(docstring: str) -> str:
 
     # Return a single string
     return '\n'.join(trimmed)
+
+
+def add_commands_to_docstring(docstring: str, plugin_docstrings: list[str]) -> str:
+    """Add commands to docstring"""
+    import sonusai
+
+    lines = docstring.splitlines()
+
+    start = lines.index('The sonusai commands are:')
+    end = lines.index('', start)
+
+    commands = sonusai.commands_doc.splitlines()
+    for plugin_docstring in plugin_docstrings:
+        commands.extend(plugin_docstring.splitlines())
+    commands.sort()
+    commands = list(filter(None, commands))
+
+    lines = lines[:start + 1] + commands + lines[end:]
+
+    return '\n'.join(lines)

sonusai/utils/model_utils.py

@@ -0,0 +1,30 @@
+from typing import Any
+
+
+def import_module(name: str) -> Any:
+    """Import a Python module adding the module file's directory to the Python system path so that relative package
+    imports are found correctly.
+    """
+    import os
+    import sys
+    from importlib import import_module
+
+    from sonusai import SonusAIError
+
+    try:
+        path = os.path.dirname(name)
+        if len(path) < 1:
+            path = './'
+
+        # Add model file location to system path
+        sys.path.append(os.path.abspath(path))
+
+        try:
+            root = os.path.splitext(os.path.basename(name))[0]
+            model = import_module(root)
+        except Exception as e:
+            raise SonusAIError(f'Error: could not import model from {name}: {e}.')
+    except Exception as e:
+        raise SonusAIError(f'Error: could not find {name}: {e}.')
+
+    return model

sonusai/utils/onnx_utils.py

@@ -3,33 +3,15 @@ from dataclasses import dataclass
 from onnxruntime import InferenceSession
 
 
-def replace_stateful_grus(keras_model, onnx_model):
-    """Replace stateful GRUs with custom layers."""
-    import warnings
-
-    with warnings.catch_warnings():
-        warnings.simplefilter('ignore')
-        from keras.layers import GRU
-
-    stateful_gru_names = []
-    for i in range(len(keras_model.layers)):
-        layer = keras_model.layers[i]
-        if isinstance(layer, GRU):
-            if layer.stateful:
-                stateful_gru_names.append(layer.name)
-
-    for node_index in range(len(onnx_model.graph.node)):
-        node = onnx_model.graph.node[node_index]
-        replace = False
-        if node.op_type == 'GRU':
-            for i in node.input:
-                for n in stateful_gru_names:
-                    if n in i:
-                        replace = True
-        if node.name in stateful_gru_names or replace:
-            node.op_type = 'SGRU'
-
-    return onnx_model
+@dataclass(frozen=True)
+class SonusAIMetaData:
+    input_shape: list[int]
+    output_shape: list[int]
+    flattened: bool
+    timestep: bool
+    channel: bool
+    mutex: bool
+    feature: str
 
 
 def add_sonusai_metadata(model,
@@ -38,13 +20,14 @@ def add_sonusai_metadata(model,
                          has_channel: bool = False,
                          is_mutex: bool = True,
                          feature: str = ''):
-    """Add SonusAI metadata to ONNX model.
-      model           keras model
-      is_flattened    model feature data is flattened
-      has_timestep    model has timestep dimension
-      has_channel     model has channel dimension
-      is_mutex        model label output is mutually exclusive
-      feature         model feature type
+    """Add SonusAI metadata to an ONNX model.
+
+    :param model: ONNX model
+    :param is_flattened: Model feature data is flattened
+    :param has_timestep: Model has timestep dimension
+    :param has_channel: Model has channel dimension
+    :param is_mutex: Model label output is mutually exclusive
+    :param feature: Model feature type
     """
     is_flattened_flag = model.metadata_props.add()
     is_flattened_flag.key = 'is_flattened'
@@ -69,18 +52,9 @@ def add_sonusai_metadata(model,
     return model
 
 
-@dataclass(frozen=True)
-class SonusAIMetaData:
-    input_shape: list[int]
-    output_shape: list[int]
-    flattened: bool
-    timestep: bool
-    channel: bool
-    mutex: bool
-    feature: str
-
-
 def get_sonusai_metadata(model: InferenceSession) -> SonusAIMetaData:
+    """Get SonusAI metadata from an ONNX model.
+    """
     m = model.get_modelmeta().custom_metadata_map
     return SonusAIMetaData(input_shape=model.get_inputs()[0].shape,
                            output_shape=model.get_outputs()[0].shape,

sonusai/utils/reshape.py

@@ -17,14 +17,14 @@ def reshape_inputs(feature: Feature,
                    timesteps: int = 0,
                    flatten: bool = False,
                    add1ch: bool = False) -> tuple[Feature, Optional[Truth]]:
-    """Check SonusAI feature and truth data and reshape feature of size [frames, strides, bands] into
+    """Check SonusAI feature and truth data and reshape feature of size [frames, strides, feature_parameters] into
     one of several options:
 
     If timesteps > 0: (i.e., for recurrent NNs):
-      no-flatten, no-channel:   [sequences, timesteps, strides, bands]      (4-dim)
-      flatten, no-channel:      [sequences, timesteps, strides*bands]       (3-dim)
-      no-flatten, add-1channel: [sequences, timesteps, strides, bands, 1]   (5-dim)
-      flatten, add-1channel:    [sequences, timesteps, strides*bands, 1]    (4-dim)
+      no-flatten, no-channel:   [sequences, timesteps, strides, feature_parameters]      (4-dim)
+      flatten, no-channel:      [sequences, timesteps, strides*feature_parameters]       (3-dim)
+      no-flatten, add-1channel: [sequences, timesteps, strides, feature_parameters, 1]   (5-dim)
+      flatten, add-1channel:    [sequences, timesteps, strides*feature_parameters, 1]    (4-dim)
 
     If batch_size is None, then do not reshape; just calculate new input shape and return.
 
@@ -40,7 +40,7 @@ def reshape_inputs(feature: Feature,
     """
     from sonusai import SonusAIError
 
-    frames, strides, bands = feature.shape
+    frames, strides, feature_parameters = feature.shape
     if truth is not None:
         truth_frames, num_classes = truth.shape
         # Double-check correctness of inputs
@@ -50,7 +50,7 @@ def reshape_inputs(feature: Feature,
         num_classes = None
 
     if flatten:
-        feature = np.reshape(feature, (frames, strides * bands))
+        feature = np.reshape(feature, (frames, strides * feature_parameters))
 
     # Reshape for Keras/TF recurrent models that require timesteps/sequence length dimension
     if timesteps > 0:
@@ -73,14 +73,14 @@ def reshape_inputs(feature: Feature,
 
         # Reshape
         if feature.ndim == 2:  # flattened input
-            # was [frames, bands*timesteps]
-            feature = np.reshape(feature, (sequences, timesteps, strides * bands))
+            # was [frames, feature_parameters*timesteps]
+            feature = np.reshape(feature, (sequences, timesteps, strides * feature_parameters))
             if truth is not None:
                 # was [frames, num_classes]
                 truth = np.reshape(truth, (sequences, timesteps, num_classes))
         elif feature.ndim == 3:  # un-flattened input
-            # was [frames, bands, timesteps]
-            feature = np.reshape(feature, (sequences, timesteps, strides, bands))
+            # was [frames, feature_parameters, timesteps]
+            feature = np.reshape(feature, (sequences, timesteps, strides, feature_parameters))
             if truth is not None:
                 # was [frames, num_classes]
                 truth = np.reshape(truth, (sequences, timesteps, num_classes))

sonusai/utils/wave.py

@@ -5,15 +5,22 @@ from sonusai.mixture.datatypes import AudioT
 def write_wav(name: str, audio: AudioT, sample_rate: int = SAMPLE_RATE) -> None:
     """ Write a simple, uncompressed WAV file.
 
-    To write multiple channels, use a 2D array of shape [samples, channels].
+    To write multiple channels, use a 2D array of shape [channels, samples].
     The bits per sample and PCM/float are determined by the data type.
 
     """
-    import numpy as np
     import torch
     import torchaudio
 
-    if audio.ndim == 1:
-        audio = np.reshape(audio, (1, audio.shape[0]))
+    data = torch.tensor(audio)
 
-    torchaudio.save(name, torch.tensor(audio), sample_rate)
+    if data.dim() == 1:
+        data = torch.reshape(data, (1, data.shape[0]))
+    if data.dim() != 2:
+        raise ValueError(f'audio must be a 1D or 2D array')
+
+    # Assuming data has more samples than channels, check if array needs to be transposed
+    if data.shape[1] < data.shape[0]:
+        data = torch.transpose(data, 0, 1)
+
+    torchaudio.save(uri=name, src=data, sample_rate=sample_rate)

{sonusai-0.15.8.dist-info → sonusai-0.16.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sonusai
-Version: 0.15.8
+Version: 0.16.0
 Summary: Framework for building deep neural network models for sound, speech, and voice AI
 Home-page: https://aaware.com
 License: GPL-3.0-only
@@ -15,50 +15,39 @@ Classifier: Programming Language :: Python :: 3.9
 Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Requires-Dist: PyYAML (>=6.0.1,<7.0.0)
-Requires-Dist: aixplain (>=0.2.6,<0.3.0)
-Requires-Dist: ctranslate2 (==4.1.0)
 Requires-Dist: dataclasses-json (>=0.6.1,<0.7.0)
-Requires-Dist: deepgram-sdk (>=3.0.0,<4.0.0)
 Requires-Dist: docopt (>=0.6.2,<0.7.0)
-Requires-Dist: faster-whisper (>=1.0.1,<2.0.0)
 Requires-Dist: h5py (>=3.11.0,<4.0.0)
 Requires-Dist: jiwer (>=3.0.3,<4.0.0)
-Requires-Dist: keras (>=3.1.1,<4.0.0)
-Requires-Dist: keras-tuner (>=1.4.7,<2.0.0)
 Requires-Dist: librosa (>=0.10.1,<0.11.0)
-Requires-Dist: lightning (>=2.2,<2.3)
 Requires-Dist: matplotlib (>=3.8.0,<4.0.0)
 Requires-Dist: onnx (>=1.14.1,<2.0.0)
 Requires-Dist: onnxruntime (>=1.16.1,<2.0.0)
 Requires-Dist: paho-mqtt (>=2.0.0,<3.0.0)
 Requires-Dist: pandas (>=2.1.1,<3.0.0)
 Requires-Dist: pesq (>=0.0.4,<0.0.5)
-Requires-Dist: pyaaware (>=1.5.3,<2.0.0)
+Requires-Dist: pyaaware (>=1.5.7,<2.0.0)
+Requires-Dist: pyaudio (>=0.2.14,<0.3.0)
 Requires-Dist: pydub (>=0.25.1,<0.26.0)
 Requires-Dist: pystoi (>=0.4.0,<0.5.0)
-Requires-Dist: python-magic (>=0.4.27,<0.5.0)
 Requires-Dist: requests (>=2.31.0,<3.0.0)
 Requires-Dist: samplerate (>=0.2.1,<0.3.0)
 Requires-Dist: soundfile (>=0.12.1,<0.13.0)
 Requires-Dist: sox (>=1.4.1,<2.0.0)
-Requires-Dist: speechrecognition (>=3.10.1,<4.0.0)
-Requires-Dist: tensorflow (>=2.15.0,<3.0.0)
-Requires-Dist: tf2onnx (>=1.15.1,<2.0.0)
 Requires-Dist: torch (>=2.2,<2.3)
 Requires-Dist: torchaudio (>=2.2,<2.3)
-Requires-Dist: torchinfo (>=1.8.0,<2.0.0)
 Requires-Dist: tqdm (>=4.66.1,<5.0.0)
 Description-Content-Type: text/x-rst
 
-Sonus AI: Framework for simplified creation of deep NN models for sound, speech, and voice AI
+SonusAI: Framework for simplified creation of deep NN models for sound, speech, and voice AI
 
-Sonus AI includes functions for pre-processing training and validation data and
+SonusAI includes functions for pre-processing training and validation data and
 creating performance metrics reports for key types of Keras models:
 - recurrent, convolutional, or a combination (i.e. RCNNs)
 - binary, multiclass single-label, multiclass multi-label, and regression
 - training with data augmentations:  noise mixing, pitch and time stretch, etc.
 
-Sonus AI python functions are used by:
- - Aaware Inc. sonusai executable:  Easily create train/validation data, run prediction, evaluate model performance
- - Keras model scripts:             User python scripts for keras model creation, training, and prediction. These can use sonusai-specific data but also some general useful utilities for trainining rnn-based models like CRNN's, DSCRNN's, etc. in Keras
+SonusAI python functions are used by:
+ - Aaware Inc. sonusai framework:   Easily create train/validation data, run prediction, evaluate model performance
+ - Keras model scripts:             User python scripts for Keras model creation, training, and prediction. These can use sonusai-specific data but also some general useful utilities for training rnn-based models like CRNN's, DSCRNN's, etc. in Keras.