"""Defines interfaces for simple inference with pretrained models
Authors:
* Aku Rouhe 2021
* Peter Plantinga 2021
* Loren Lugosch 2020
* Mirco Ravanelli 2020
* Titouan Parcollet 2021
"""
import torch
import torchaudio
from types import SimpleNamespace
from torch.nn import SyncBatchNorm
from torch.nn import DataParallel as DP
from hyperpyyaml import load_hyperpyyaml
from speechbrain.pretrained.fetching import fetch
from speechbrain.dataio.preprocess import AudioNormalizer
import torch.nn.functional as F
from torch.nn.parallel import DistributedDataParallel as DDP
from speechbrain.utils.data_utils import split_path
from speechbrain.utils.distributed import run_on_main
[docs]class Pretrained:
"""Takes a trained model and makes predictions on new data.
This is a base class which handles some common boilerplate.
It intentionally has an interface similar to ``Brain`` - these base
classes handle similar things.
Subclasses of Pretrained should implement the actual logic of how
the pretrained system runs, and add methods with descriptive names
(e.g. transcribe_file() for ASR).
Arguments
---------
modules : dict of str:torch.nn.Module pairs
The Torch modules that make up the learned system. These can be treated
in special ways (put on the right device, frozen, etc.)
hparams : dict
Each key:value pair should consist of a string key and a hyperparameter
that is used within the overridden methods. These will
be accessible via an ``hparams`` attribute, using "dot" notation:
e.g., self.hparams.model(x).
run_opts : dict
Options parsed from command line. See ``speechbrain.parse_arguments()``.
List that are supported here:
* device
* data_parallel_count
* data_parallel_backend
* distributed_launch
* distributed_backend
* jit_module_keys
freeze_params : bool
To freeze (requires_grad=False) parameters or not. Normally in inference
you want to freeze the params. Also calls .eval() on all modules.
"""
HPARAMS_NEEDED = []
MODULES_NEEDED = []
def __init__(
self, modules=None, hparams=None, run_opts=None, freeze_params=True
):
# Arguments passed via the run opts dictionary. Set a limited
# number of these, since some don't apply to inference.
run_opt_defaults = {
"device": "cpu",
"data_parallel_count": -1,
"data_parallel_backend": False,
"distributed_launch": False,
"distributed_backend": "nccl",
"jit_module_keys": None,
}
for arg, default in run_opt_defaults.items():
if run_opts is not None and arg in run_opts:
setattr(self, arg, run_opts[arg])
else:
# If any arg from run_opt_defaults exist in hparams and
# not in command line args "run_opts"
if hparams is not None and arg in hparams:
setattr(self, arg, hparams[arg])
else:
setattr(self, arg, default)
# Put modules on the right device, accessible with dot notation
self.modules = torch.nn.ModuleDict(modules)
for mod in self.modules:
self.modules[mod].to(self.device)
for mod in self.MODULES_NEEDED:
if mod not in modules:
raise ValueError(f"Need modules['{mod}']")
# Check MODULES_NEEDED and HPARAMS_NEEDED and
# make hyperparams available with dot notation
if self.HPARAMS_NEEDED and hparams is None:
raise ValueError("Need to provide hparams dict.")
if hparams is not None:
# Also first check that all required params are found:
for hp in self.HPARAMS_NEEDED:
if hp not in hparams:
raise ValueError(f"Need hparams['{hp}']")
self.hparams = SimpleNamespace(**hparams)
# Prepare modules for computation, e.g. jit
self._prepare_modules(freeze_params)
# Audio normalization
self.audio_normalizer = hparams.get(
"audio_normalizer", AudioNormalizer()
)
def _prepare_modules(self, freeze_params):
"""Prepare modules for computation, e.g. jit.
Arguments
---------
freeze_params : bool
Whether to freeze the parameters and call ``eval()``.
"""
# Make jit-able
self._compile_jit()
self._wrap_distributed()
# If we don't want to backprop, freeze the pretrained parameters
if freeze_params:
self.modules.eval()
for p in self.modules.parameters():
p.requires_grad = False
[docs] def load_audio(self, path, savedir="."):
"""Load an audio file with this model"s input spec
When using a speech model, it is important to use the same type of data,
as was used to train the model. This means for example using the same
sampling rate and number of channels. It is, however, possible to
convert a file from a higher sampling rate to a lower one (downsampling).
Similarly, it is simple to downmix a stereo file to mono.
The path can be a local path, a web url, or a link to a huggingface repo.
"""
source, fl = split_path(path)
path = fetch(fl, source=source, savedir=savedir)
signal, sr = torchaudio.load(path, channels_first=False)
return self.audio_normalizer(signal, sr)
def _compile_jit(self):
"""Compile requested modules with ``torch.jit.script``."""
if self.jit_module_keys is None:
return
for name in self.jit_module_keys:
if name not in self.modules:
raise ValueError(
"module " + name + " cannot be jit compiled because "
"it is not defined in your hparams file."
)
module = torch.jit.script(self.modules[name])
self.modules[name] = module.to(self.device)
def _wrap_distributed(self):
"""Wrap modules with distributed wrapper when requested."""
if not self.distributed_launch and not self.data_parallel_backend:
return
elif self.distributed_launch:
for name, module in self.modules.items():
if any(p.requires_grad for p in module.parameters()):
# for ddp, all module must run on same GPU
module = SyncBatchNorm.convert_sync_batchnorm(module)
module = DDP(module, device_ids=[self.device])
self.modules[name] = module
else:
# data_parallel_backend
for name, module in self.modules.items():
if any(p.requires_grad for p in module.parameters()):
# if distributed_count = -1 then use all gpus
# otherwise, specify the set of gpu to use
if self.data_parallel_count == -1:
module = DP(module)
else:
module = DP(
module, [i for i in range(self.data_parallel_count)]
)
self.modules[name] = module
[docs] @classmethod
def from_hparams(
cls,
source,
hparams_file="hyperparams.yaml",
overrides={},
savedir=None,
use_auth_token=False,
**kwargs,
):
"""Fetch and load based from outside source based on HyperPyYAML file
The source can be a location on the filesystem or online/huggingface
The hyperparams file should contain a "modules" key, which is a
dictionary of torch modules used for computation.
The hyperparams file should contain a "pretrainer" key, which is a
speechbrain.utils.parameter_transfer.Pretrainer
Arguments
---------
source : str
The location to use for finding the model. See
``speechbrain.pretrained.fetching.fetch`` for details.
hparams_file : str
The name of the hyperparameters file to use for constructing
the modules necessary for inference. Must contain two keys:
"modules" and "pretrainer", as described.
overrides : dict
Any changes to make to the hparams file when it is loaded.
savedir : str or Path
Where to put the pretraining material. If not given, will use
./pretrained_models/<class-name>-hash(source).
use_auth_token : bool (default: False)
If true Hugginface's auth_token will be used to load private models from the HuggingFace Hub,
default is False because majority of models are public.
"""
if savedir is None:
clsname = cls.__name__
savedir = f"./pretrained_models/{clsname}-{hash(source)}"
hparams_local_path = fetch(
hparams_file, source, savedir, use_auth_token
)
# Load the modules:
with open(hparams_local_path) as fin:
hparams = load_hyperpyyaml(fin, overrides)
# Pretraining:
pretrainer = hparams["pretrainer"]
pretrainer.set_collect_in(savedir)
# For distributed setups, have this here:
run_on_main(pretrainer.collect_files, kwargs={"default_source": source})
# Load on the CPU. Later the params can be moved elsewhere by specifying
# run_opts={"device": ...}
pretrainer.load_collected(device="cpu")
# Now return the system
return cls(hparams["modules"], hparams, **kwargs)
[docs]class EndToEndSLU(Pretrained):
"""A end-to-end SLU model.
The class can be used either to run only the encoder (encode()) to extract
features or to run the entire model (decode()) to map the speech to its semantics.
Example
-------
>>> from speechbrain.pretrained import EndToEndSLU
>>> tmpdir = getfixture("tmpdir")
>>> slu_model = EndToEndSLU.from_hparams(
... source="speechbrain/slu-timers-and-such-direct-librispeech-asr",
... savedir=tmpdir,
... )
>>> slu_model.decode_file("samples/audio_samples/example6.wav")
"{'intent': 'SimpleMath', 'slots': {'number1': 37.67, 'number2': 75.7, 'op': ' minus '}}"
"""
HPARAMS_NEEDED = ["tokenizer", "asr_model_source"]
MODULES_NEEDED = ["slu_enc", "beam_searcher"]
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.tokenizer = self.hparams.tokenizer
self.asr_model = EncoderDecoderASR.from_hparams(
source=self.hparams.asr_model_source,
run_opts={"device": self.device},
)
[docs] def decode_file(self, path):
"""Maps the given audio file to a string representing the
semantic dictionary for the utterance.
Arguments
---------
path : str
Path to audio file to decode.
Returns
-------
str
The predicted semantics.
"""
waveform = self.load_audio(path)
waveform = waveform.to(self.device)
# Fake a batch:
batch = waveform.unsqueeze(0)
rel_length = torch.tensor([1.0])
predicted_words, predicted_tokens = self.decode_batch(batch, rel_length)
return predicted_words[0]
[docs] def encode_batch(self, wavs, wav_lens):
"""Encodes the input audio into a sequence of hidden states
Arguments
---------
wavs : torch.tensor
Batch of waveforms [batch, time, channels] or [batch, time]
depending on the model.
wav_lens : torch.tensor
Lengths of the waveforms relative to the longest one in the
batch, tensor of shape [batch]. The longest one should have
relative length 1.0 and others len(waveform) / max_length.
Used for ignoring padding.
Returns
-------
torch.tensor
The encoded batch
"""
wavs = wavs.float()
wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
with torch.no_grad():
ASR_encoder_out = self.asr_model.encode_batch(
wavs.detach(), wav_lens
)
encoder_out = self.modules.slu_enc(ASR_encoder_out)
return encoder_out
[docs] def decode_batch(self, wavs, wav_lens):
"""Maps the input audio to its semantics
Arguments
---------
wavs : torch.tensor
Batch of waveforms [batch, time, channels] or [batch, time]
depending on the model.
wav_lens : torch.tensor
Lengths of the waveforms relative to the longest one in the
batch, tensor of shape [batch]. The longest one should have
relative length 1.0 and others len(waveform) / max_length.
Used for ignoring padding.
Returns
-------
list
Each waveform in the batch decoded.
tensor
Each predicted token id.
"""
with torch.no_grad():
wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
encoder_out = self.encode_batch(wavs, wav_lens)
predicted_tokens, scores = self.modules.beam_searcher(
encoder_out, wav_lens
)
predicted_words = [
self.tokenizer.decode_ids(token_seq)
for token_seq in predicted_tokens
]
return predicted_words, predicted_tokens
[docs]class EncoderDecoderASR(Pretrained):
"""A ready-to-use Encoder-Decoder ASR model
The class can be used either to run only the encoder (encode()) to extract
features or to run the entire encoder-decoder model
(transcribe()) to transcribe speech. The given YAML must contains the fields
specified in the *_NEEDED[] lists.
Example
-------
>>> from speechbrain.pretrained import EncoderDecoderASR
>>> tmpdir = getfixture("tmpdir")
>>> asr_model = EncoderDecoderASR.from_hparams(
... source="speechbrain/asr-crdnn-rnnlm-librispeech",
... savedir=tmpdir,
... )
>>> asr_model.transcribe_file("samples/audio_samples/example2.flac")
"MY FATHER HAS REVEALED THE CULPRIT'S NAME"
"""
HPARAMS_NEEDED = ["tokenizer"]
MODULES_NEEDED = ["encoder", "decoder"]
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.tokenizer = self.hparams.tokenizer
[docs] def transcribe_file(self, path):
"""Transcribes the given audiofile into a sequence of words.
Arguments
---------
path : str
Path to audio file which to transcribe.
Returns
-------
str
The audiofile transcription produced by this ASR system.
"""
waveform = self.load_audio(path)
# Fake a batch:
batch = waveform.unsqueeze(0)
rel_length = torch.tensor([1.0])
predicted_words, predicted_tokens = self.transcribe_batch(
batch, rel_length
)
return predicted_words[0]
[docs] def encode_batch(self, wavs, wav_lens):
"""Encodes the input audio into a sequence of hidden states
The waveforms should already be in the model's desired format.
You can call:
``normalized = EncoderDecoderASR.normalizer(signal, sample_rate)``
to get a correctly converted signal in most cases.
Arguments
---------
wavs : torch.tensor
Batch of waveforms [batch, time, channels] or [batch, time]
depending on the model.
wav_lens : torch.tensor
Lengths of the waveforms relative to the longest one in the
batch, tensor of shape [batch]. The longest one should have
relative length 1.0 and others len(waveform) / max_length.
Used for ignoring padding.
Returns
-------
torch.tensor
The encoded batch
"""
wavs = wavs.float()
wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
encoder_out = self.modules.encoder(wavs, wav_lens)
return encoder_out
[docs] def transcribe_batch(self, wavs, wav_lens):
"""Transcribes the input audio into a sequence of words
The waveforms should already be in the model's desired format.
You can call:
``normalized = EncoderDecoderASR.normalizer(signal, sample_rate)``
to get a correctly converted signal in most cases.
Arguments
---------
wavs : torch.tensor
Batch of waveforms [batch, time, channels] or [batch, time]
depending on the model.
wav_lens : torch.tensor
Lengths of the waveforms relative to the longest one in the
batch, tensor of shape [batch]. The longest one should have
relative length 1.0 and others len(waveform) / max_length.
Used for ignoring padding.
Returns
-------
list
Each waveform in the batch transcribed.
tensor
Each predicted token id.
"""
with torch.no_grad():
wav_lens = wav_lens.to(self.device)
encoder_out = self.encode_batch(wavs, wav_lens)
predicted_tokens, scores = self.modules.decoder(
encoder_out, wav_lens
)
predicted_words = [
self.tokenizer.decode_ids(token_seq)
for token_seq in predicted_tokens
]
return predicted_words, predicted_tokens
[docs]class EncoderASR(Pretrained):
"""A ready-to-use Encoder ASR model
The class can be used either to run only the encoder (encode()) to extract
features or to run the entire encoder + decoder function model
(transcribe()) to transcribe speech. The given YAML must contains the fields
specified in the *_NEEDED[] lists.
Example
-------
>>> from speechbrain.pretrained import EncoderASR
>>> tmpdir = getfixture("tmpdir")
>>> asr_model = EncoderASR.from_hparams(
... source="speechbrain/asr-wav2vec2-commonvoice-fr",
... savedir=tmpdir,
... ) # doctest: +SKIP
>>> asr_model.transcribe_file("samples/audio_samples/example_fr.wav") # doctest: +SKIP
"""
HPARAMS_NEEDED = ["tokenizer", "decoding_function"]
MODULES_NEEDED = ["encoder"]
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.tokenizer = self.hparams.tokenizer
self.decoding_function = self.hparams.decoding_function
[docs] def transcribe_file(self, path):
"""Transcribes the given audiofile into a sequence of words.
Arguments
---------
path : str
Path to audio file which to transcribe.
Returns
-------
str
The audiofile transcription produced by this ASR system.
"""
waveform = self.load_audio(path)
# Fake a batch:
batch = waveform.unsqueeze(0)
rel_length = torch.tensor([1.0])
predicted_words, predicted_tokens = self.transcribe_batch(
batch, rel_length
)
return str(predicted_words[0])
[docs] def encode_batch(self, wavs, wav_lens):
"""Encodes the input audio into a sequence of hidden states
The waveforms should already be in the model's desired format.
You can call:
``normalized = EncoderASR.normalizer(signal, sample_rate)``
to get a correctly converted signal in most cases.
Arguments
---------
wavs : torch.tensor
Batch of waveforms [batch, time, channels] or [batch, time]
depending on the model.
wav_lens : torch.tensor
Lengths of the waveforms relative to the longest one in the
batch, tensor of shape [batch]. The longest one should have
relative length 1.0 and others len(waveform) / max_length.
Used for ignoring padding.
Returns
-------
torch.tensor
The encoded batch
"""
wavs = wavs.float()
wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
encoder_out = self.modules.encoder(wavs, wav_lens)
return encoder_out
[docs] def transcribe_batch(self, wavs, wav_lens):
"""Transcribes the input audio into a sequence of words
The waveforms should already be in the model's desired format.
You can call:
``normalized = EncoderASR.normalizer(signal, sample_rate)``
to get a correctly converted signal in most cases.
Arguments
---------
wavs : torch.tensor
Batch of waveforms [batch, time, channels] or [batch, time]
depending on the model.
wav_lens : torch.tensor
Lengths of the waveforms relative to the longest one in the
batch, tensor of shape [batch]. The longest one should have
relative length 1.0 and others len(waveform) / max_length.
Used for ignoring padding.
Returns
-------
list
Each waveform in the batch transcribed.
tensor
Each predicted token id.
"""
with torch.no_grad():
wav_lens = wav_lens.to(self.device)
encoder_out = self.encode_batch(wavs, wav_lens)
predictions = self.decoding_function(encoder_out, wav_lens)
predicted_words = [
self.tokenizer.decode_ids(token_seq)
for token_seq in predictions
]
return predicted_words, predictions
[docs]class EncoderClassifier(Pretrained):
"""A ready-to-use class for utterance-level classification (e.g, speaker-id,
language-id, emotion recognition, keyword spotting, etc).
The class assumes that an encoder called "embedding_model" and a model
called "classifier" are defined in the yaml file. If you want to
convert the predicted index into a corresponding text label, please
provide the path of the label_encoder in a variable called 'lab_encoder_file'
within the yaml.
The class can be used either to run only the encoder (encode_batch()) to
extract embeddings or to run a classification step (classify_batch()).
```
Example
-------
>>> import torchaudio
>>> from speechbrain.pretrained import EncoderClassifier
>>> # Model is downloaded from the speechbrain HuggingFace repo
>>> tmpdir = getfixture("tmpdir")
>>> classifier = EncoderClassifier.from_hparams(
... source="speechbrain/spkrec-ecapa-voxceleb",
... savedir=tmpdir,
... )
>>> # Compute embeddings
>>> signal, fs = torchaudio.load("samples/audio_samples/example1.wav")
>>> embeddings = classifier.encode_batch(signal)
>>> # Classification
>>> prediction = classifier .classify_batch(signal)
"""
MODULES_NEEDED = [
"compute_features",
"mean_var_norm",
"embedding_model",
"classifier",
]
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
[docs] def encode_batch(self, wavs, wav_lens=None, normalize=False):
"""Encodes the input audio into a single vector embedding.
The waveforms should already be in the model's desired format.
You can call:
``normalized = <this>.normalizer(signal, sample_rate)``
to get a correctly converted signal in most cases.
Arguments
---------
wavs : torch.tensor
Batch of waveforms [batch, time, channels] or [batch, time]
depending on the model. Make sure the sample rate is fs=16000 Hz.
wav_lens : torch.tensor
Lengths of the waveforms relative to the longest one in the
batch, tensor of shape [batch]. The longest one should have
relative length 1.0 and others len(waveform) / max_length.
Used for ignoring padding.
normalize : bool
If True, it normalizes the embeddings with the statistics
contained in mean_var_norm_emb.
Returns
-------
torch.tensor
The encoded batch
"""
# Manage single waveforms in input
if len(wavs.shape) == 1:
wavs = wavs.unsqueeze(0)
# Assign full length if wav_lens is not assigned
if wav_lens is None:
wav_lens = torch.ones(wavs.shape[0], device=self.device)
# Storing waveform in the specified device
wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
wavs = wavs.float()
# Computing features and embeddings
feats = self.modules.compute_features(wavs)
feats = self.modules.mean_var_norm(feats, wav_lens)
embeddings = self.modules.embedding_model(feats, wav_lens)
if normalize:
embeddings = self.hparams.mean_var_norm_emb(
embeddings, torch.ones(embeddings.shape[0], device=self.device)
)
return embeddings
[docs] def classify_batch(self, wavs, wav_lens=None):
"""Performs classification on the top of the encoded features.
It returns the posterior probabilities, the index and, if the label
encoder is specified it also the text label.
Arguments
---------
wavs : torch.tensor
Batch of waveforms [batch, time, channels] or [batch, time]
depending on the model. Make sure the sample rate is fs=16000 Hz.
wav_lens : torch.tensor
Lengths of the waveforms relative to the longest one in the
batch, tensor of shape [batch]. The longest one should have
relative length 1.0 and others len(waveform) / max_length.
Used for ignoring padding.
Returns
-------
out_prob
The log posterior probabilities of each class ([batch, N_class])
score:
It is the value of the log-posterior for the best class ([batch,])
index
The indexes of the best class ([batch,])
text_lab:
List with the text labels corresponding to the indexes.
(label encoder should be provided).
"""
emb = self.encode_batch(wavs, wav_lens)
out_prob = self.modules.classifier(emb).squeeze(1)
score, index = torch.max(out_prob, dim=-1)
text_lab = self.hparams.label_encoder.decode_torch(index)
return out_prob, score, index, text_lab
[docs] def classify_file(self, path):
"""Classifies the given audiofile into the given set of labels.
Arguments
---------
path : str
Path to audio file to classify.
Returns
-------
out_prob
The log posterior probabilities of each class ([batch, N_class])
score:
It is the value of the log-posterior for the best class ([batch,])
index
The indexes of the best class ([batch,])
text_lab:
List with the text labels corresponding to the indexes.
(label encoder should be provided).
"""
waveform = self.load_audio(path)
# Fake a batch:
batch = waveform.unsqueeze(0)
rel_length = torch.tensor([1.0])
emb = self.encode_batch(batch, rel_length)
out_prob = self.modules.classifier(emb).squeeze(1)
score, index = torch.max(out_prob, dim=-1)
text_lab = self.hparams.label_encoder.decode_torch(index)
return out_prob, score, index, text_lab
[docs]class SpeakerRecognition(EncoderClassifier):
"""A ready-to-use model for speaker recognition. It can be used to
perform speaker verification with verify_batch().
```
Example
-------
>>> import torchaudio
>>> from speechbrain.pretrained import SpeakerRecognition
>>> # Model is downloaded from the speechbrain HuggingFace repo
>>> tmpdir = getfixture("tmpdir")
>>> verification = SpeakerRecognition.from_hparams(
... source="speechbrain/spkrec-ecapa-voxceleb",
... savedir=tmpdir,
... )
>>> # Perform verification
>>> signal, fs = torchaudio.load("samples/audio_samples/example1.wav")
>>> signal2, fs = torchaudio.load("samples/audio_samples/example2.flac")
>>> score, prediction = verification.verify_batch(signal, signal2)
"""
MODULES_NEEDED = [
"compute_features",
"mean_var_norm",
"embedding_model",
"mean_var_norm_emb",
]
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.similarity = torch.nn.CosineSimilarity(dim=-1, eps=1e-6)
[docs] def verify_batch(
self, wavs1, wavs2, wav1_lens=None, wav2_lens=None, threshold=0.25
):
"""Performs speaker verification with cosine distance.
It returns the score and the decision (0 different speakers,
1 same speakers).
Arguments
---------
wavs1 : Torch.Tensor
Tensor containing the speech waveform1 (batch, time).
Make sure the sample rate is fs=16000 Hz.
wavs2 : Torch.Tensor
Tensor containing the speech waveform2 (batch, time).
Make sure the sample rate is fs=16000 Hz.
wav1_lens: Torch.Tensor
Tensor containing the relative length for each sentence
in the length (e.g., [0.8 0.6 1.0])
wav2_lens: Torch.Tensor
Tensor containing the relative length for each sentence
in the length (e.g., [0.8 0.6 1.0])
threshold: Float
Threshold applied to the cosine distance to decide if the
speaker is different (0) or the same (1).
Returns
-------
score
The score associated to the binary verification output
(cosine distance).
prediction
The prediction is 1 if the two signals in input are from the same
speaker and 0 otherwise.
"""
emb1 = self.encode_batch(wavs1, wav1_lens, normalize=True)
emb2 = self.encode_batch(wavs2, wav2_lens, normalize=True)
score = self.similarity(emb1, emb2)
return score, score > threshold
[docs] def verify_files(self, path_x, path_y, threshold=0.25):
"""Speaker verification with cosine distance
Returns the score and the decision (0 different speakers,
1 same speakers).
Returns
-------
score
The score associated to the binary verification output
(cosine distance).
prediction
The prediction is 1 if the two signals in input are from the same
speaker and 0 otherwise.
"""
waveform_x = self.load_audio(path_x)
waveform_y = self.load_audio(path_y)
# Fake batches:
batch_x = waveform_x.unsqueeze(0)
batch_y = waveform_y.unsqueeze(0)
# Verify:
score, decision = self.verify_batch(
batch_x, batch_y, threshold=threshold
)
# Squeeze:
return score[0], decision[0]
[docs]class VAD(Pretrained):
"""A ready-to-use class for Voice Activity Detection (VAD) using a
pre-trained model.
Example
-------
>>> import torchaudio
>>> from speechbrain.pretrained import VAD
>>> # Model is downloaded from the speechbrain HuggingFace repo
>>> tmpdir = getfixture("tmpdir")
>>> VAD = VAD.from_hparams(
... source="speechbrain/vad-crdnn-libriparty",
... savedir=tmpdir,
... )
>>> # Perform VAD
>>> boundaries = VAD.get_speech_segments("samples/audio_samples/example1.wav")
"""
HPARAMS_NEEDED = ["sample_rate", "time_resolution", "device"]
MODULES_NEEDED = ["compute_features", "mean_var_norm", "model"]
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.time_resolution = self.hparams.time_resolution
self.sample_rate = self.hparams.sample_rate
self.device = self.hparams.device
[docs] def get_speech_prob_file(
self,
audio_file,
large_chunk_size=30,
small_chunk_size=10,
overlap_small_chunk=False,
):
"""Outputs the frame-level speech probability of the input audio file
using the neural model specified in the hparam file. To make this code
both parallelizable and scalable to long sequences, it uses a
double-windowing approach. First, we sequentially read non-overlapping
large chunks of the input signal. We then split the large chunks into
smaller chunks and we process them in parallel.
Arguments
---------
audio_file: path
Path of the audio file containing the recording. The file is read
with torchaudio.
large_chunk_size: float
Size (in seconds) of the large chunks that are read sequentially
from the input audio file.
small_chunk_size:
Size (in seconds) of the small chunks extracted from the large ones.
The audio signal is processed in parallel within the small chunks.
Note that large_chunk_size/small_chunk_size must be an integer.
overlap_small_chunk: bool
True, creates overlapped small chunks. The probabilities of the
overlapped chunks are combined using hamming windows.
Returns
-------
prob_vad: torch.tensor
Tensor containing the frame-level speech probabilities for the
input audio file.
"""
# Getting the total size of the input file
sample_rate, audio_len = self._get_audio_info(audio_file)
if sample_rate != self.sample_rate:
raise ValueError(
"The detected sample rate is different from that set in the hparam file"
)
# Computing the length (in samples) of the large and small chunks
long_chunk_len = int(sample_rate * large_chunk_size)
small_chunk_len = int(sample_rate * small_chunk_size)
# Setting the step size of the small chunk (50% overapping windows are supported)
small_chunk_step = small_chunk_size
if overlap_small_chunk:
small_chunk_step = small_chunk_size / 2
# Computing the length (in sample) of the small_chunk step size
small_chunk_len_step = int(sample_rate * small_chunk_step)
# Loop over big chunks
prob_chunks = []
last_chunk = False
begin_sample = 0
while True:
# Reading the big chunk
large_chunk, fs = torchaudio.load(
audio_file, frame_offset=begin_sample, num_frames=long_chunk_len
)
large_chunk = large_chunk.to(self.device)
# Manage padding of the last small chunk
if last_chunk or large_chunk.shape[-1] < small_chunk_len:
padding = torch.zeros(
1, small_chunk_len, device=large_chunk.device
)
large_chunk = torch.cat([large_chunk, padding], dim=1)
# Splitting the big chunk into smaller (overlapped) ones
small_chunks = torch.nn.functional.unfold(
large_chunk.unsqueeze(1).unsqueeze(2),
kernel_size=(1, small_chunk_len),
stride=(1, small_chunk_len_step),
)
small_chunks = small_chunks.squeeze(0).transpose(0, 1)
# Getting (in parallel) the frame-level speech probabilities
small_chunks_prob = self.get_speech_prob_chunk(small_chunks)
small_chunks_prob = small_chunks_prob[:, :-1, :]
# Manage overlapping chunks
if overlap_small_chunk:
small_chunks_prob = self._manage_overlapped_chunks(
small_chunks_prob
)
# Prepare for folding
small_chunks_prob = small_chunks_prob.permute(2, 1, 0)
# Computing lengths in samples
out_len = int(
large_chunk.shape[-1] / (sample_rate * self.time_resolution)
)
kernel_len = int(small_chunk_size / self.time_resolution)
step_len = int(small_chunk_step / self.time_resolution)
# Folding the frame-level predictions
small_chunks_prob = torch.nn.functional.fold(
small_chunks_prob,
output_size=(1, out_len),
kernel_size=(1, kernel_len),
stride=(1, step_len),
)
# Appending the frame-level speech probabilities of the large chunk
small_chunks_prob = small_chunks_prob.squeeze(1).transpose(-1, -2)
prob_chunks.append(small_chunks_prob)
# Check stop condition
if last_chunk:
break
# Update counter to process the next big chunk
begin_sample = begin_sample + long_chunk_len
# Check if the current chunk is the last one
if begin_sample + long_chunk_len > audio_len:
last_chunk = True
# Converting the list to a tensor
prob_vad = torch.cat(prob_chunks, dim=1)
last_elem = int(audio_len / (self.time_resolution * sample_rate))
prob_vad = prob_vad[:, 0:last_elem, :]
return prob_vad
def _manage_overlapped_chunks(self, small_chunks_prob):
"""This support function manages overlapped the case in which the
small chunks have a 50% overlap."""
# Weighting the frame-level probabilities with a hamming window
# reduces uncertainnty when overlapping chunks are used.
hamming_window = torch.hamming_window(
small_chunks_prob.shape[1], device=self.device
)
# First and last chunks require special care
half_point = int(small_chunks_prob.shape[1] / 2)
small_chunks_prob[0, half_point:] = small_chunks_prob[
0, half_point:
] * hamming_window[half_point:].unsqueeze(1)
small_chunks_prob[-1, 0:half_point] = small_chunks_prob[
-1, 0:half_point
] * hamming_window[0:half_point].unsqueeze(1)
# Applying the window to all the other probabilities
small_chunks_prob[1:-1] = small_chunks_prob[
1:-1
] * hamming_window.unsqueeze(0).unsqueeze(2)
return small_chunks_prob
[docs] def get_speech_prob_chunk(self, wavs, wav_lens=None):
"""Outputs the frame-level posterior probability for the input audio chunks
Outputs close to zero refers to time steps with a low probability of speech
activity, while outputs closer to one likely contain speech.
Arguments
---------
wavs : torch.tensor
Batch of waveforms [batch, time, channels] or [batch, time]
depending on the model. Make sure the sample rate is fs=16000 Hz.
wav_lens : torch.tensor
Lengths of the waveforms relative to the longest one in the
batch, tensor of shape [batch]. The longest one should have
relative length 1.0 and others len(waveform) / max_length.
Used for ignoring padding.
Returns
-------
torch.tensor
The encoded batch
"""
# Manage single waveforms in input
if len(wavs.shape) == 1:
wavs = wavs.unsqueeze(0)
# Assign full length if wav_lens is not assigned
if wav_lens is None:
wav_lens = torch.ones(wavs.shape[0], device=self.device)
# Storing waveform in the specified device
wavs, wav_lens = wavs.to(self.device), wav_lens.to(self.device)
wavs = wavs.float()
# Computing features and embeddings
feats = self.modules.compute_features(wavs)
feats = self.modules.mean_var_norm(feats, wav_lens)
outputs = self.modules.cnn(feats)
outputs = outputs.reshape(
outputs.shape[0],
outputs.shape[1],
outputs.shape[2] * outputs.shape[3],
)
outputs, h = self.modules.rnn(outputs)
outputs = self.modules.dnn(outputs)
output_prob = torch.sigmoid(outputs)
return output_prob
[docs] def apply_threshold(
self, vad_prob, activation_th=0.5, deactivation_th=0.25
):
"""Scans the frame-level speech probabilities and applies a threshold
on them. Speech starts when a value larger than activation_th is
detected, while it ends when observing a value lower than
the deactivation_th.
Arguments
---------
vad_prob: torch.tensor
Frame-level speech probabilities.
activation_th: float
Threshold for starting a speech segment.
deactivation_th: float
Threshold for ending a speech segment.
Returns
-------
vad_th: torch.tensor
Tensor containing 1 for speech regions and 0 for non-speech regions.
"""
vad_activation = (vad_prob >= activation_th).int()
vad_deactivation = (vad_prob >= deactivation_th).int()
vad_th = vad_activation + vad_deactivation
# Loop over batches and time steps
for batch in range(vad_th.shape[0]):
for time_step in range(vad_th.shape[1] - 1):
if (
vad_th[batch, time_step] == 2
and vad_th[batch, time_step + 1] == 1
):
vad_th[batch, time_step + 1] = 2
vad_th[vad_th == 1] = 0
vad_th[vad_th == 2] = 1
return vad_th
[docs] def get_boundaries(self, prob_th, output_value="seconds"):
"""Computes the time boundaries where speech activity is detected.
It takes in input frame-level binary decisions
(1 for speech, 0 for non-speech) and outputs the begin/end second
(or sample) of each detected speech region.
Arguments
---------
prob_th: torch.tensor
Frame-level binary decisions (1 for speech frame, 0 for a
non-speech one). The tensor can be obtained from apply_threshold.
put_value: 'seconds' or 'samples'
When the option 'seconds' is set, the returned boundaries are in
seconds, otherwise, it reports them in samples.
Returns
-------
boundaries: torch.tensor
Tensor containing the start second (or sample) of speech segments
in even positions and their corresponding end in odd positions
(e.g, [1.0, 1.5, 5,.0 6.0] means that we have two speech segment;
one from 1.0 to 1.5 seconds and another from 5.0 to 6.0 seconds).
"""
# Shifting frame-levels binary decision by 1
# This allows detecting changes in speech/non-speech activities
prob_th_shifted = torch.roll(prob_th, dims=1, shifts=1)
prob_th_shifted[:, 0, :] = 0
prob_th = prob_th + prob_th_shifted
# Needed to manage last time step
prob_th[:, -1, :] = (prob_th[:, -1, :] >= 1).int()
# Where prob_th is 1 there is a change
indexes = (prob_th == 1).nonzero()[:, 1].reshape(-1, 2)
# Remove 1 from end samples
indexes[:, -1] = indexes[:, -1] - 1
# From indexes to samples
seconds = (indexes * self.time_resolution).float()
samples = (self.sample_rate * seconds).int()
if output_value == "seconds":
boundaries = seconds
else:
boundaries = samples
return boundaries
[docs] def merge_short_segments(self, boundaries, close_th=0.250):
"""Merges segments that are shorter than the given threshold.
Arguments
---------
boundaries : str
Tensor containing the speech boundaries. It can be derived using the
get_boundaries method.
close_th: float
If the distance between boundaries is smaller than close_th, the
segments will be merged.
Returns
-------
new_boundaries
The new boundaries with the merged segments.
"""
new_boudaries = []
# Single segment case
if boundaries.shape[0] == 0:
return boundaries
# Getting beg and end of previous segment
prev_beg_seg = boundaries[0, 0].float()
prev_end_seg = boundaries[0, 1].float()
# Process all the segments
for i in range(1, boundaries.shape[0]):
beg_seg = boundaries[i, 0]
segment_distance = beg_seg - prev_end_seg
# Mergin close segments
if segment_distance <= close_th:
prev_end_seg = boundaries[i, 1]
else:
# Appending new segments
new_boudaries.append([prev_beg_seg, prev_end_seg])
prev_beg_seg = beg_seg
prev_end_seg = boundaries[i, 1]
new_boudaries.append([prev_beg_seg, prev_end_seg])
new_boudaries = torch.FloatTensor(new_boudaries).to(boundaries.device)
return new_boudaries
[docs] def remove_short_segments(self, boundaries, len_th=0.250):
"""Removes segments that are too short.
Arguments
---------
boundaries : str
Tensor containing the speech boundaries. It can be derived using the
get_boundaries method.
len_th: float
If the length of the segment is smaller than close_th, the segments
will be merged.
Returns
-------
new_boundaries
The new boundaries without the short segments.
"""
new_boundaries = []
# Process the segments
for i in range(boundaries.shape[0]):
# Computing segment length
seg_len = boundaries[i, 1] - boundaries[i, 0]
# Accept segment only if longer than len_th
if seg_len > len_th:
new_boundaries.append([boundaries[i, 0], boundaries[i, 1]])
new_boundaries = torch.FloatTensor(new_boundaries).to(boundaries.device)
return new_boundaries
[docs] def save_boundaries(
self, boundaries, save_path=None, print_boundaries=True
):
"""Saves the boundaries on a file (and/or prints them) in a readable format.
Arguments
---------
boundaries: torch.tensor
Tensor containing the speech boundaries. It can be derived using the
get_boundaries method.
save_path: path
When to store the text file containing the speech/non-speech intervals.
print_boundaries: Bool
Prints the speech/non-speech intervals in the standard outputs.
"""
# Create a new file if needed
if save_path is not None:
f = open(save_path, mode="w", encoding="utf-8")
# Setting the rights format for second- or sample-based boundaries
if boundaries.dtype == torch.int:
value_format = "% i"
else:
value_format = "% .2f "
# Printing speech and non-speech intervals
last_end = 0
cnt_seg = 0
for i in range(boundaries.shape[0]):
begin_value = boundaries[i, 0]
end_value = boundaries[i, 1]
if last_end != begin_value:
cnt_seg = cnt_seg + 1
print_str = (
"segment_%03d " + value_format + value_format + "NON_SPEECH"
)
if print_boundaries:
print(print_str % (cnt_seg, last_end, begin_value))
if save_path is not None:
f.write(print_str % (cnt_seg, last_end, begin_value) + "\n")
cnt_seg = cnt_seg + 1
print_str = "segment_%03d " + value_format + value_format + "SPEECH"
if print_boundaries:
print(print_str % (cnt_seg, begin_value, end_value))
if save_path is not None:
f.write(print_str % (cnt_seg, last_end, begin_value) + "\n")
last_end = end_value
if save_path is not None:
f.close()
[docs] def energy_VAD(
self, audio_file, boundaries, activation_th=0.5, deactivation_th=0.0
):
"""Applies energy-based VAD within the detected speech segments.The neural
network VAD often creates longer segments and tends to merge segments that
are close with each other.
The energy VAD post-processes can be useful for having a fine-grained voice
activity detection.
The energy VAD computes the energy within the small chunks. The energy is
normalized within the segment to have mean 0.5 and +-0.5 of std.
This helps to set the energy threshold.
Arguments
---------
audio_file: path
Path of the audio file containing the recording. The file is read
with torchaudio.
boundaries : str
Tensor containing the speech boundaries. It can be derived using the
get_boundaries method.
activation_th: float
A new speech segment is started it the energy is above activation_th.
deactivation_th: float
The segment is considered ended when the energy is <= deactivation_th.
Returns
-------
new_boundaries
The new boundaries that are post-processed by the energy VAD.
"""
# Getting the total size of the input file
sample_rate, audio_len = self._get_audio_info(audio_file)
if sample_rate != self.sample_rate:
raise ValueError(
"The detected sample rate is different from that set in the hparam file"
)
# Computing the chunk length of the energy window
chunk_len = int(self.time_resolution * sample_rate)
new_boundaries = []
# Processing speech segments
for i in range(boundaries.shape[0]):
begin_sample = int(boundaries[i, 0] * sample_rate)
end_sample = int(boundaries[i, 1] * sample_rate)
seg_len = end_sample - begin_sample
# Reading the speech segment
segment, _ = torchaudio.load(
audio_file, frame_offset=begin_sample, num_frames=seg_len
)
# Create chunks
segment_chunks = self.create_chunks(
segment, chunk_size=chunk_len, chunk_stride=chunk_len
)
# Energy computtion within each chunk
energy_chunks = segment_chunks.abs().sum(-1).log()
# Energy normalization
energy_chunks = (
(energy_chunks - energy_chunks.mean())
/ (2 * energy_chunks.std())
) + 0.5
energy_chunks = energy_chunks.unsqueeze(0).unsqueeze(2)
# Apply threshold based on the energy value
energy_vad = self.apply_threshold(
energy_chunks,
activation_th=activation_th,
deactivation_th=deactivation_th,
)
# Get the boundaries
energy_boundaries = self.get_boundaries(
energy_vad, output_value="seconds"
)
# Get the final boundaries in the original signal
for j in range(energy_boundaries.shape[0]):
start_en = boundaries[i, 0] + energy_boundaries[j, 0]
end_end = boundaries[i, 0] + energy_boundaries[j, 1]
new_boundaries.append([start_en, end_end])
# Convert boundaries to tensor
new_boundaries = torch.FloatTensor(new_boundaries).to(boundaries.device)
return new_boundaries
[docs] def create_chunks(self, x, chunk_size=16384, chunk_stride=16384):
"""Splits the input into smaller chunks of size chunk_size with
an overlap chunk_stride. The chunks are concatenated over
the batch axis.
Arguments
---------
x: torch.Tensor
Signal to split into chunks.
chunk_size : str
The size of each chunk.
chunk_stride:
The stride (hop) of each chunk.
Returns
-------
x: torch.Tensor
A new tensors with the chunks derived from the input signal.
"""
x = x.unfold(1, chunk_size, chunk_stride)
x = x.reshape(x.shape[0] * x.shape[1], -1)
return x
def _get_audio_info(self, audio_file):
"""Returns the sample rate and the length of the input audio file"""
# Getting the total size of the input file
metadata = torchaudio.info(audio_file)
sample_rate = metadata.sample_rate
audio_len = metadata.num_frames
return sample_rate, audio_len
[docs] def upsample_VAD(self, vad_out, audio_file, time_resolution=0.01):
"""Upsamples the output of the vad to help visualization. It creates a
signal that is 1 when there is speech and 0 when there is no speech.
The vad signal has the same resolution as the input one and can be
opened with it (e.g, using audacity) to visually figure out VAD regions.
Arguments
---------
vad_out: torch.Tensor
Tensor containing 1 for each frame of speech and 0 for each non-speech
frame.
audio_file: path
The original audio file used to compute vad_out
time_resolution : float
Time resolution of the vad_out signal.
Returns
-------
vad_signal
The upsampled version of the vad_out tensor.
"""
# Getting the total size of the input file
sample_rate, sig_len = self._get_audio_info(audio_file)
if sample_rate != self.sample_rate:
raise ValueError(
"The detected sample rate is different from that set in the hparam file"
)
beg_samp = 0
step_size = int(time_resolution * sample_rate)
end_samp = step_size
index = 0
# Initialize upsampled signal
vad_signal = torch.zeros(1, sig_len, device=vad_out.device)
# Upsample signal
while end_samp < sig_len:
vad_signal[0, beg_samp:end_samp] = vad_out[0, index, 0]
index = index + 1
beg_samp = beg_samp + step_size
end_samp = beg_samp + step_size
return vad_signal
[docs] def upsample_boundaries(self, boundaries, audio_file):
"""Based on the input boundaries, this method creates a signal that is 1
when there is speech and 0 when there is no speech.
The vad signal has the same resolution as the input one and can be
opened with it (e.g, using audacity) to visually figure out VAD regions.
Arguments
---------
boundaries: torch.Tensor
Tensor containing the boundaries of the speech segments.
audio_file: path
The original audio file used to compute vad_out
Returns
-------
vad_signal
The output vad signal with the same resolution of the input one.
"""
# Getting the total size of the input file
sample_rate, sig_len = self._get_audio_info(audio_file)
if sample_rate != self.sample_rate:
raise ValueError(
"The detected sample rate is different from that set in the hparam file"
)
# Initialization of the output signal
vad_signal = torch.zeros(1, sig_len, device=boundaries.device)
# Composing the vad signal from boundaries
for i in range(boundaries.shape[0]):
beg_sample = int(boundaries[i, 0] * sample_rate)
end_sample = int(boundaries[i, 1] * sample_rate)
vad_signal[0, beg_sample:end_sample] = 1.0
return vad_signal
[docs] def double_check_speech_segments(
self, boundaries, audio_file, speech_th=0.5
):
"""Takes in input the boundaries of the detected speech segments and
double checks (using the neural VAD) that they actually contain speech.
Arguments
---------
boundaries: torch.Tensor
Tensor containing the boundaries of the speech segments.
audio_file: path
The original audio file used to compute vad_out.
speech_th: float
Threshold on the mean posterior probability over which speech is
confirmed. Below that threshold, the segment is re-assigned to a
non-speech region.
Returns
-------
new_boundaries
The boundaries of the segments where speech activity is confirmed.
"""
# Getting the total size of the input file
sample_rate, sig_len = self._get_audio_info(audio_file)
# Double check the segments
new_boundaries = []
for i in range(boundaries.shape[0]):
beg_sample = int(boundaries[i, 0] * sample_rate)
end_sample = int(boundaries[i, 1] * sample_rate)
len_seg = end_sample - beg_sample
# Read the candidate speech segment
segment, fs = torchaudio.load(
audio_file, frame_offset=beg_sample, num_frames=len_seg
)
speech_prob = self.get_speech_prob_chunk(segment)
if speech_prob.mean() > speech_th:
# Accept this a as a speech segment
new_boundaries.append([boundaries[i, 0], boundaries[i, 1]])
# Convert boundaries from list to tensor
new_boundaries = torch.FloatTensor(new_boundaries).to(boundaries.device)
return new_boundaries
[docs] def get_segments(
self, boundaries, audio_file, before_margin=0.1, after_margin=0.1
):
"""Returns a list containing all the detected speech segments.
Arguments
---------
boundaries: torch.Tensor
Tensor containing the boundaries of the speech segments.
audio_file: path
The original audio file used to compute vad_out.
before_margin: float
Used to cut the segments samples a bit before the detected margin.
after_margin: float
Use to cut the segments samples a bit after the detected margin.
Returns
-------
segments: list
List containing the detected speech segments
"""
sample_rate, sig_len = self._get_audio_info(audio_file)
if sample_rate != self.sample_rate:
raise ValueError(
"The detected sample rate is different from that set in the hparam file"
)
segments = []
for i in range(boundaries.shape[0]):
beg_sample = boundaries[i, 0] * sample_rate
end_sample = boundaries[i, 1] * sample_rate
beg_sample = int(max(0, beg_sample - before_margin * sample_rate))
end_sample = int(
min(sig_len, end_sample + after_margin * sample_rate)
)
len_seg = end_sample - beg_sample
vad_segment, fs = torchaudio.load(
audio_file, frame_offset=beg_sample, num_frames=len_seg
)
segments.append(vad_segment)
return segments
[docs] def get_speech_segments(
self,
audio_file,
large_chunk_size=30,
small_chunk_size=10,
overlap_small_chunk=False,
apply_energy_VAD=False,
double_check=True,
close_th=0.250,
len_th=0.250,
activation_th=0.5,
deactivation_th=0.25,
en_activation_th=0.5,
en_deactivation_th=0.0,
speech_th=0.50,
):
"""Detects speech segments within the input file. The input signal can
be both a short or a long recording. The function computes the
posterior probabilities on large chunks (e.g, 30 sec), that are read
sequentially (to avoid storing big signals in memory).
Each large chunk is, in turn, split into smaller chunks (e.g, 10 seconds)
that are processed in parallel. The pipeline for detecting the speech
segments is the following:
1- Compute posteriors probabilities at the frame level.
2- Apply a threshold on the posterior probability.
3- Derive candidate speech segments on top of that.
4- Apply energy VAD within each candidate segment (optional).
5- Merge segments that are too close.
6- Remove segments that are too short.
7- Double check speech segments (optional).
Arguments
---------
audio_file : str
Path to audio file.
large_chunk_size: float
Size (in seconds) of the large chunks that are read sequentially
from the input audio file.
small_chunk_size: float
Size (in seconds) of the small chunks extracted from the large ones.
The audio signal is processed in parallel within the small chunks.
Note that large_chunk_size/small_chunk_size must be an integer.
overlap_small_chunk: bool
If True, it creates overlapped small chunks (with 50% overal).
The probabilities of the overlapped chunks are combined using
hamming windows.
apply_energy_VAD: bool
If True, a energy-based VAD is used on the detected speech segments.
The neural network VAD often creates longer segments and tends to
merge close segments together. The energy VAD post-processes can be
useful for having a fine-grained voice activity detection.
The energy thresholds is managed by activation_th and
deactivation_th (see below).
double_check: bool
If True, double checkis (using the neural VAD) that the candidate
speech segments actually contain speech. A threshold on the mean
posterior probabilities provided by the neural network is applied
based on the speech_th parameter (see below).
activation_th: float
Threshold of the neural posteriors above which starting a speech segment.
deactivation_th: float
Threshold of the neural posteriors below which ending a speech segment.
en_activation_th: float
A new speech segment is started it the energy is above activation_th.
This is active only if apply_energy_VAD is True.
en_deactivation_th: float
The segment is considered ended when the energy is <= deactivation_th.
This is active only if apply_energy_VAD is True.
speech_th: float
Threshold on the mean posterior probability within the candidate
speech segment. Below that threshold, the segment is re-assigned to
a non-speech region. This is active only if double_check is True.
close_th: float
If the distance between boundaries is smaller than close_th, the
segments will be merged.
len_th: float
If the length of the segment is smaller than close_th, the segments
will be merged.
Returns
-------
boundaries: torch.tensor
Tensor containing the start second of speech segments in even
positions and their corresponding end in odd positions
(e.g, [1.0, 1.5, 5,.0 6.0] means that we have two speech segment;
one from 1.0 to 1.5 seconds and another from 5.0 to 6.0 seconds).
"""
# Fetch audio file from web if not local
source, fl = split_path(audio_file)
audio_file = fetch(fl, source=source)
# Computing speech vs non speech probabilities
prob_chunks = self.get_speech_prob_file(
audio_file,
large_chunk_size=large_chunk_size,
small_chunk_size=small_chunk_size,
overlap_small_chunk=overlap_small_chunk,
)
# Apply a threshold to get candidate speech segments
prob_th = self.apply_threshold(
prob_chunks,
activation_th=activation_th,
deactivation_th=deactivation_th,
).float()
# Comupute the boundaries of the speech segments
boundaries = self.get_boundaries(prob_th, output_value="seconds")
# Apply energy-based VAD on the detected speech segments
if apply_energy_VAD:
boundaries = self.energy_VAD(
audio_file,
boundaries,
activation_th=en_activation_th,
deactivation_th=en_deactivation_th,
)
# Merge short segments
boundaries = self.merge_short_segments(boundaries, close_th=close_th)
# Remove short segments
boundaries = self.remove_short_segments(boundaries, len_th=len_th)
# Double check speech segments
if double_check:
boundaries = self.double_check_speech_segments(
boundaries, audio_file, speech_th=speech_th
)
return boundaries
[docs]class SpectralMaskEnhancement(Pretrained):
"""A ready-to-use model for speech enhancement.
Arguments
---------
See ``Pretrained``.
Example
-------
>>> import torchaudio
>>> from speechbrain.pretrained import SpectralMaskEnhancement
>>> # Model is downloaded from the speechbrain HuggingFace repo
>>> tmpdir = getfixture("tmpdir")
>>> enhancer = SpectralMaskEnhancement.from_hparams(
... source="speechbrain/mtl-mimic-voicebank",
... savedir=tmpdir,
... )
>>> noisy, fs = torchaudio.load("samples/audio_samples/example_noisy.wav")
>>> # Channel dimension is interpreted as batch dimension here
>>> enhanced = enhancer.enhance_batch(noisy)
"""
HPARAMS_NEEDED = ["compute_stft", "spectral_magnitude", "resynth"]
MODULES_NEEDED = ["enhance_model"]
[docs] def compute_features(self, wavs):
"""Compute the log spectral magnitude features for masking.
Arguments
---------
wavs : torch.tensor
A batch of waveforms to convert to log spectral mags.
"""
feats = self.hparams.compute_stft(wavs)
feats = self.hparams.spectral_magnitude(feats)
return torch.log1p(feats)
[docs] def enhance_batch(self, noisy, lengths=None):
"""Enhance a batch of noisy waveforms.
Arguments
---------
noisy : torch.tensor
A batch of waveforms to perform enhancement on.
lengths : torch.tensor
The lengths of the waveforms if the enhancement model handles them.
Returns
-------
torch.tensor
A batch of enhanced waveforms of the same shape as input.
"""
noisy = noisy.to(self.device)
noisy_features = self.compute_features(noisy)
# Perform masking-based enhancement, multiplying output with input.
if lengths is not None:
mask = self.modules.enhance_model(noisy_features, lengths=lengths)
else:
mask = self.modules.enhance_model(noisy_features)
enhanced = torch.mul(mask, noisy_features)
# Return resynthesized waveforms
return self.hparams.resynth(torch.expm1(enhanced), noisy)
[docs] def enhance_file(self, filename, output_filename=None):
"""Enhance a wav file.
Arguments
---------
filename : str
Location on disk to load file for enhancement.
output_filename : str
If provided, writes enhanced data to this file.
"""
noisy = self.load_audio(filename)
noisy = noisy.to(self.device)
# Fake a batch:
batch = noisy.unsqueeze(0)
enhanced = self.enhance_batch(batch)
if output_filename is not None:
torchaudio.save(output_filename, enhanced, channels_first=False)
return enhanced.squeeze(0)