speechbrain.pretrained.interfaces module

Defines interfaces for simple inference with pretrained models

Authors:

Aku Rouhe 2021
Peter Plantinga 2021
Loren Lugosch 2020
Mirco Ravanelli 2020
Titouan Parcollet 2021
Abdel Heba 2021
Andreas Nautsch 2022

Summary

Classes:

`EncodeDecodePipelineMixin`	A mixin for pretrained models that makes it possible to specify an encoding pipeline and a decoding pipeline
`EncoderASR`	A ready-to-use Encoder ASR model
`EncoderClassifier`	A ready-to-use class for utterance-level classification (e.g, speaker-id, language-id, emotion recognition, keyword spotting, etc).
`EncoderDecoderASR`	A ready-to-use Encoder-Decoder ASR model
`EndToEndSLU`	A end-to-end SLU model.
`GraphemeToPhoneme`	A pretrained model implementation for Grapheme-to-Phoneme (G2P) models that take raw natural language text as an input and
`HIFIGAN`	A ready-to-use wrapper for HiFiGAN (mel_spec -> waveform).
`Pretrained`	Takes a trained model and makes predictions on new data.
`SNREstimator`	A "ready-to-use" SNR estimator.
`SepformerSeparation`	A "ready-to-use" speech separation model.
`SpeakerRecognition`	A ready-to-use model for speaker recognition.
`SpectralMaskEnhancement`	A ready-to-use model for speech enhancement.
`Tacotron2`	A ready-to-use wrapper for Tacotron2 (text -> mel_spec).
`VAD`	A ready-to-use class for Voice Activity Detection (VAD) using a pre-trained model.
`WaveformEncoder`	A ready-to-use waveformEncoder model
`WaveformEnhancement`	A ready-to-use model for speech enhancement.

Functions:

foreign_class

Fetch and load an interface from an outside source

Reference

speechbrain.pretrained.interfaces.foreign_class(source, hparams_file='hyperparams.yaml', pymodule_file='custom.py', classname='CustomInterface', overrides={}, savedir=None, use_auth_token=False, download_only=False, **kwargs)[source]

Fetch and load an interface from an outside source

The source can be a location on the filesystem or online/huggingface

The pymodule file should contain a class with the given classname. An instance of that class is returned. The idea is to have a custom Pretrained subclass in the file. The pymodule file is also added to the python path before the Hyperparams YAML file is loaded, so it can contain any custom implementations that are needed.

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

Parameters

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.
pymodule_file (str) – The name of the Python file that should be fetched.
classname (str) – The name of the Class, of which an instance is created and returned
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.
download_only (bool (default: False)) – If true, class and instance creation is skipped.

Returns

An instance of a class with the given classname from the given pymodule file.

Return type

object

class speechbrain.pretrained.interfaces.Pretrained(modules=None, hparams=None, run_opts=None, freeze_params=True)[source]

Bases: Module

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).

Pretrained is a torch.nn.Module so that methods like .to() or .eval() can work. Subclasses should provide a suitable forward() implementation: by convention, it should be a method that takes a batch of audio signals and runs the full model (as applicable).

Parameters

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.). These are available as attributes under self.mods, like self.mods.model(x)
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 = []

load_audio(path, savedir='.')[source]

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.

classmethod from_hparams(source, hparams_file='hyperparams.yaml', pymodule_file='custom.py', overrides={}, savedir=None, use_auth_token=False, revision=None, download_only=False, **kwargs)[source]

Fetch and load based from outside source based on HyperPyYAML file

The source can be a location on the filesystem or online/huggingface

You can use the pymodule_file to include any custom implementations that are needed: if that file exists, then its location is added to sys.path before Hyperparams YAML is loaded, so it can be referenced in the YAML.

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

Parameters

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.
pymodule_file (str) – A Python file can be fetched. This allows any custom implementations to be included. The file’s location is added to sys.path before the hyperparams YAML file is loaded, so it can be referenced in YAML. This is optional, but has a default: “custom.py”. If the default file is not found, this is simply ignored, but if you give a different filename, then this will raise in case the file is not found.
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.
revision (str) – The model revision corresponding to the HuggingFace Hub model revision. This is particularly useful if you wish to pin your code to a particular version of a model hosted at HuggingFace.
download_only (bool (default: False)) – If true, class and instance creation is skipped.

training: bool

class speechbrain.pretrained.interfaces.EndToEndSLU(*args, **kwargs)[source]

Bases: 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("tests/samples/single-mic/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']

decode_file(path)[source]

Maps the given audio file to a string representing the semantic dictionary for the utterance.

Parameters: path (str) – Path to audio file to decode.
Returns: The predicted semantics.
Return type: str

encode_batch(wavs, wav_lens)[source]

Encodes the input audio into a sequence of hidden states

Parameters

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

The encoded batch

Return type

torch.tensor

decode_batch(wavs, wav_lens)[source]

Maps the input audio to its semantics

Parameters

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.

forward(wavs, wav_lens)[source]: Runs full decoding - note: no gradients through decoding

training: bool

class speechbrain.pretrained.interfaces.EncoderDecoderASR(*args, **kwargs)[source]

Bases: 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("tests/samples/single-mic/example2.flac")
"MY FATHER HAS REVEALED THE CULPRIT'S NAME"

HPARAMS_NEEDED = ['tokenizer']

MODULES_NEEDED = ['encoder', 'decoder']

transcribe_file(path)[source]

Transcribes the given audiofile into a sequence of words.

Parameters: path (str) – Path to audio file which to transcribe.
Returns: The audiofile transcription produced by this ASR system.
Return type: str

encode_batch(wavs, wav_lens)[source]

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.

Parameters

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

The encoded batch

Return type

torch.tensor

transcribe_batch(wavs, wav_lens)[source]

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.

Parameters

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.

forward(wavs, wav_lens)[source]: Runs full transcription - note: no gradients through decoding

training: bool

class speechbrain.pretrained.interfaces.WaveformEncoder(*args, **kwargs)[source]

Bases: Pretrained

A ready-to-use waveformEncoder model

It can be used to wrap different embedding models such as SSL ones (wav2vec2) or speaker ones (Xvector) etc. Two functions are available: encode_batch and encode_file. They can be used to obtain the embeddings directly from an audio file or from a batch of audio tensors respectively.

The given YAML must contains the fields specified in the *_NEEDED[] lists.

Example

>>> from speechbrain.pretrained import WaveformEncoder
>>> tmpdir = getfixture("tmpdir")
>>> ssl_model = WaveformEncoder.from_hparams(
...     source="speechbrain/ssl-wav2vec2-base-libri",
...     savedir=tmpdir,
... ) 
>>> ssl_model.encode_file("samples/audio_samples/example_fr.wav") 

MODULES_NEEDED = ['encoder']

encode_file(path)[source]

Encode the given audiofile into a sequence of embeddings.

Parameters: path (str) – Path to audio file which to encode.
Returns: The audiofile embeddings produced by this system.
Return type: torch.tensor

encode_batch(wavs, wav_lens)[source]

Encodes the input audio into a sequence of hidden states

The waveforms should already be in the model’s desired format.

Parameters

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

The encoded batch

Return type

torch.tensor

forward(wavs, wav_lens)[source]: Runs the encoder

training: bool

class speechbrain.pretrained.interfaces.EncoderASR(*args, **kwargs)[source]

Bases: 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,
... ) 
>>> asr_model.transcribe_file("samples/audio_samples/example_fr.wav") 

HPARAMS_NEEDED = ['tokenizer', 'decoding_function']

MODULES_NEEDED = ['encoder']

transcribe_file(path)[source]

Transcribes the given audiofile into a sequence of words.

Parameters: path (str) – Path to audio file which to transcribe.
Returns: The audiofile transcription produced by this ASR system.
Return type: str

encode_batch(wavs, wav_lens)[source]

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.

Parameters

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

The encoded batch

Return type

torch.tensor

transcribe_batch(wavs, wav_lens)[source]

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.

Parameters

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.

forward(wavs, wav_lens)[source]: Runs the encoder

training: bool

class speechbrain.pretrained.interfaces.EncoderClassifier(*args, **kwargs)[source]

Bases: 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("tests/samples/single-mic/example1.wav")
>>> embeddings = classifier.encode_batch(signal)

>>> # Classification
>>> prediction = classifier.classify_batch(signal)

MODULES_NEEDED = ['compute_features', 'mean_var_norm', 'embedding_model', 'classifier']

encode_batch(wavs, wav_lens=None, normalize=False)[source]

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.

Parameters

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

The encoded batch

Return type

torch.tensor

classify_batch(wavs, wav_lens=None)[source]

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.

Parameters

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).

classify_file(path)[source]

Classifies the given audiofile into the given set of labels.

Parameters

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).

forward(wavs, wav_lens=None)[source]: Runs the classification

training: bool

class speechbrain.pretrained.interfaces.SpeakerRecognition(*args, **kwargs)[source]

Bases: EncoderClassifier

A ready-to-use model for speaker recognition. It can be used to perform speaker verification with verify_batch().

``` .. rubric:: 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("tests/samples/single-mic/example1.wav")
>>> signal2, fs = torchaudio.load("tests/samples/single-mic/example2.flac")
>>> score, prediction = verification.verify_batch(signal, signal2)

MODULES_NEEDED = ['compute_features', 'mean_var_norm', 'embedding_model', 'mean_var_norm_emb']

verify_batch(wavs1, wavs2, wav1_lens=None, wav2_lens=None, threshold=0.25)[source]

Performs speaker verification with cosine distance.

It returns the score and the decision (0 different speakers, 1 same speakers).

Parameters

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.

verify_files(path_x, path_y)[source]

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.

training: bool

class speechbrain.pretrained.interfaces.VAD(*args, **kwargs)[source]

Bases: 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("tests/samples/single-mic/example1.wav")

HPARAMS_NEEDED = ['sample_rate', 'time_resolution', 'device']

MODULES_NEEDED = ['compute_features', 'mean_var_norm', 'model']

get_speech_prob_file(audio_file, large_chunk_size=30, small_chunk_size=10, overlap_small_chunk=False)[source]

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.

Parameters

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 – Tensor containing the frame-level speech probabilities for the input audio file.

Return type

torch.tensor

get_speech_prob_chunk(wavs, wav_lens=None)[source]

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.

Parameters

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

The encoded batch

Return type

torch.tensor

apply_threshold(vad_prob, activation_th=0.5, deactivation_th=0.25)[source]

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.

Parameters

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 – Tensor containing 1 for speech regions and 0 for non-speech regions.

Return type

torch.tensor

get_boundaries(prob_th, output_value='seconds')[source]

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.

Parameters

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 – 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).

Return type

torch.tensor

merge_close_segments(boundaries, close_th=0.25)[source]

Merges segments that are shorter than the given threshold.

Parameters

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

The new boundaries with the merged segments.

Return type

new_boundaries

remove_short_segments(boundaries, len_th=0.25)[source]

Removes segments that are too short.

Parameters

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

The new boundaries without the short segments.

Return type

new_boundaries

save_boundaries(boundaries, save_path=None, print_boundaries=True, audio_file=None)[source]

Saves the boundaries on a file (and/or prints them) in a readable format.

Parameters

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.
audio_file (path) – Path of the audio file containing the recording. The file is read with torchaudio. It is used here to detect the length of the signal.

energy_VAD(audio_file, boundaries, activation_th=0.5, deactivation_th=0.0, eps=1e-06)[source]

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.

Parameters

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.
eps (float) – Small constant for numerical stability.

Returns

The new boundaries that are post-processed by the energy VAD.

Return type

new_boundaries

create_chunks(x, chunk_size=16384, chunk_stride=16384)[source]

Splits the input into smaller chunks of size chunk_size with an overlap chunk_stride. The chunks are concatenated over the batch axis.

Parameters

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 – A new tensors with the chunks derived from the input signal.

Return type

torch.Tensor

upsample_VAD(vad_out, audio_file, time_resolution=0.01)[source]

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.

Parameters

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

The upsampled version of the vad_out tensor.

Return type

vad_signal

upsample_boundaries(boundaries, audio_file)[source]

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.

Parameters

boundaries (torch.Tensor) – Tensor containing the boundaries of the speech segments.
audio_file (path) – The original audio file used to compute vad_out

Returns

The output vad signal with the same resolution of the input one.

Return type

vad_signal

double_check_speech_segments(boundaries, audio_file, speech_th=0.5)[source]

Takes in input the boundaries of the detected speech segments and double checks (using the neural VAD) that they actually contain speech.

Parameters

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

The boundaries of the segments where speech activity is confirmed.

Return type

new_boundaries

get_segments(boundaries, audio_file, before_margin=0.1, after_margin=0.1)[source]

Returns a list containing all the detected speech segments.

Parameters

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 containing the detected speech segments

Return type

list

get_speech_segments(audio_file, large_chunk_size=30, small_chunk_size=10, overlap_small_chunk=False, apply_energy_VAD=False, double_check=True, close_th=0.25, len_th=0.25, activation_th=0.5, deactivation_th=0.25, en_activation_th=0.5, en_deactivation_th=0.0, speech_th=0.5)[source]

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).

Parameters

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 – 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).

Return type

torch.tensor

forward(wavs, wav_lens=None)[source]: Gets frame-level speech-activity predictions

training: bool

class speechbrain.pretrained.interfaces.SepformerSeparation(modules=None, hparams=None, run_opts=None, freeze_params=True)[source]

Bases: Pretrained

A “ready-to-use” speech separation model.

Uses Sepformer architecture.

Example

>>> tmpdir = getfixture("tmpdir")
>>> model = SepformerSeparation.from_hparams(
...     source="speechbrain/sepformer-wsj02mix",
...     savedir=tmpdir)
>>> mix = torch.randn(1, 400)
>>> est_sources = model.separate_batch(mix)
>>> print(est_sources.shape)
torch.Size([1, 400, 2])

MODULES_NEEDED = ['encoder', 'masknet', 'decoder']

separate_batch(mix)[source]

Run source separation on batch of audio.

Parameters: mix (torch.tensor) – The mixture of sources.
Returns: Separated sources
Return type: tensor

separate_file(path, savedir='.')[source]

Separate sources from file.

Parameters

path (str) – Path to file which has a mixture of sources. It can be a local path, a web url, or a huggingface repo.
savedir (path) – Path where to store the wav signals (when downloaded from the web).

Returns

Separated sources

Return type

tensor

forward(mix)[source]: Runs separation on the input mix

training: bool

class speechbrain.pretrained.interfaces.SpectralMaskEnhancement(modules=None, hparams=None, run_opts=None, freeze_params=True)[source]

Bases: Pretrained

A ready-to-use model for speech enhancement.

Parameters: Pretrained. (See) –

Example

>>> import torch
>>> from speechbrain.pretrained import SpectralMaskEnhancement
>>> # Model is downloaded from the speechbrain HuggingFace repo
>>> tmpdir = getfixture("tmpdir")
>>> enhancer = SpectralMaskEnhancement.from_hparams(
...     source="speechbrain/metricgan-plus-voicebank",
...     savedir=tmpdir,
... )
>>> enhanced = enhancer.enhance_file(
...     "speechbrain/metricgan-plus-voicebank/example.wav"
... )

HPARAMS_NEEDED = ['compute_stft', 'spectral_magnitude', 'resynth']

MODULES_NEEDED = ['enhance_model']

compute_features(wavs)[source]

Compute the log spectral magnitude features for masking.

Parameters: wavs (torch.tensor) – A batch of waveforms to convert to log spectral mags.

enhance_batch(noisy, lengths=None)[source]

Enhance a batch of noisy waveforms.

Parameters

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

A batch of enhanced waveforms of the same shape as input.

Return type

torch.tensor

enhance_file(filename, output_filename=None)[source]

Enhance a wav file.

Parameters

filename (str) – Location on disk to load file for enhancement.
output_filename (str) – If provided, writes enhanced data to this file.

training: bool

class speechbrain.pretrained.interfaces.EncodeDecodePipelineMixin[source]

Bases: object

A mixin for pretrained models that makes it possible to specify an encoding pipeline and a decoding pipeline

create_pipelines()[source]: Initializes the encode and decode pipeline

to_dict(data)[source]

Converts padded batches to dictionaries, leaves other data types as is

Parameters: data (object) – a dictionary or a padded batch
Returns: results – the dictionary
Return type: dict

property batch_inputs

Determines whether the input pipeline operates on batches or individual examples (true means batched)

Returns: batch_intputs
Return type: bool

property input_use_padded_data

If turned on, raw PaddedData instances will be passed to the model. If turned off, only .data will be used

Returns: result – whether padded data is used as is
Return type: bool

property batch_outputs

Determines whether the output pipeline operates on batches or individual examples (true means batched)

Returns: batch_outputs
Return type: bool

encode_input(input)[source]

Encodes the inputs using the pipeline

Parameters

input (dict) – the raw inputs
Results –
------- –
results (object) –

decode_output(output)[source]

Decodes the raw model outputs

Parameters

output (tuple) – raw model outputs
Results –
------- –
result (dict or list) – the output of the pipeline

class speechbrain.pretrained.interfaces.GraphemeToPhoneme(*args, **kwargs)[source]

Bases: Pretrained, EncodeDecodePipelineMixin

A pretrained model implementation for Grapheme-to-Phoneme (G2P) models that take raw natural language text as an input and

Example

>>> text = ("English is tough. It can be understood "
...         "through thorough thought though")
>>> from speechbrain.pretrained import GraphemeToPhoneme
>>> tmpdir = getfixture('tmpdir')
>>> g2p = GraphemeToPhoneme.from_hparams('path/to/model', savedir=tmpdir) 
>>> phonemes = g2p.g2p(text) 

INPUT_STATIC_KEYS = ['txt']

OUTPUT_KEYS = ['phonemes']

property phonemes: Returns the available phonemes

property language: Returns the language for which this model is available

g2p(text)[source]

Performs the Grapheme-to-Phoneme conversion

Parameters: text (str or list[str]) – a single string to be encoded to phonemes - or a sequence of strings
Returns: result – if a single example was provided, the return value is a single list of phonemes
Return type: list

load_dependencies()[source]: Loads any relevant model dependencies

__call__(text)[source]

A convenience callable wrapper - same as G2P

Parameters: text (str or list[str]) – a single string to be encoded to phonemes - or a sequence of strings
Returns: result – if a single example was provided, the return value is a single list of phonemes
Return type: list

forward(noisy, lengths=None)[source]: Runs enhancement on the noisy input

training: bool

class speechbrain.pretrained.interfaces.WaveformEnhancement(modules=None, hparams=None, run_opts=None, freeze_params=True)[source]

Bases: Pretrained

A ready-to-use model for speech enhancement.

Parameters: Pretrained. (See) –

Example

>>> from speechbrain.pretrained import WaveformEnhancement
>>> # Model is downloaded from the speechbrain HuggingFace repo
>>> tmpdir = getfixture("tmpdir")
>>> enhancer = WaveformEnhancement.from_hparams(
...     source="speechbrain/mtl-mimic-voicebank",
...     savedir=tmpdir,
... )
>>> enhanced = enhancer.enhance_file(
...     "speechbrain/mtl-mimic-voicebank/example.wav"
... )

MODULES_NEEDED = ['enhance_model']

enhance_batch(noisy, lengths=None)[source]

Enhance a batch of noisy waveforms.

Parameters

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

A batch of enhanced waveforms of the same shape as input.

Return type

torch.tensor

enhance_file(filename, output_filename=None)[source]

Enhance a wav file.

Parameters

filename (str) – Location on disk to load file for enhancement.
output_filename (str) – If provided, writes enhanced data to this file.

forward(noisy, lengths=None)[source]: Runs enhancement on the noisy input

training: bool

class speechbrain.pretrained.interfaces.SNREstimator(modules=None, hparams=None, run_opts=None, freeze_params=True)[source]

Bases: Pretrained

A “ready-to-use” SNR estimator.

MODULES_NEEDED = ['encoder', 'encoder_out']

HPARAMS_NEEDED = ['stat_pooling', 'snrmax', 'snrmin']

estimate_batch(mix, predictions)[source]

Run SI-SNR estimation on the estimated sources, and mixture.

Parameters

mix (torch.tensor) – The mixture of sources of shape B X T
predictions (torch.tensor) –
of size (B x T x C), where B is batch size

T is number of time points C is number of sources

Returns

Estimate of SNR

Return type

tensor

forward(mix, predictions)[source]: Just run the batch estimate

gettrue_snrrange(inp)[source]: Convert from 0-1 range to true snr range

training: bool

class speechbrain.pretrained.interfaces.Tacotron2(*args, **kwargs)[source]

Bases: Pretrained

A ready-to-use wrapper for Tacotron2 (text -> mel_spec).

Parameters: hparams – Hyperparameters (from HyperPyYAML)

Example

>>> tmpdir_vocoder = getfixture('tmpdir') / "vocoder"
>>> tacotron2 = Tacotron2.from_hparams(source="speechbrain/tts-tacotron2-ljspeech", savedir=tmpdir_vocoder)
>>> mel_output, mel_length, alignment = tacotron2.encode_text("Mary had a little lamb")
>>> items = [
...   "A quick brown fox jumped over the lazy dog",
...   "How much wood would a woodchuck chuck?",
...   "Never odd or even"
... ]
>>> mel_outputs, mel_lengths, alignments = tacotron2.encode_batch(items)

>>> # One can combine the TTS model with a vocoder (that generates the final waveform)
>>> # Intialize the Vocoder (HiFIGAN)
>>> tmpdir_tts = getfixture('tmpdir') / "tts"
>>> hifi_gan = HIFIGAN.from_hparams(source="speechbrain/tts-hifigan-ljspeech", savedir=tmpdir_tts)
>>> # Running the TTS
>>> mel_output, mel_length, alignment = tacotron2.encode_text("Mary had a little lamb")
>>> # Running Vocoder (spectrogram-to-waveform)
>>> waveforms = hifi_gan.decode_batch(mel_output)

HPARAMS_NEEDED = ['model', 'text_to_sequence']

text_to_seq(txt)[source]: Encodes raw text into a tensor with a customer text-to-equence fuction

encode_batch(texts)[source]

Computes mel-spectrogram for a list of texts

Texts must be sorted in decreasing order on their lengths

Parameters: text (List[str]) – texts to be encoded into spectrogram
Return type: tensors of output spectrograms, output lengths and alignments

encode_text(text)[source]: Runs inference for a single text str

forward(texts)[source]: Encodes the input texts.

training: bool

class speechbrain.pretrained.interfaces.HIFIGAN(*args, **kwargs)[source]

Bases: Pretrained

A ready-to-use wrapper for HiFiGAN (mel_spec -> waveform).

Parameters: hparams – Hyperparameters (from HyperPyYAML)

Example

>>> tmpdir_vocoder = getfixture('tmpdir') / "vocoder"
>>> hifi_gan = HIFIGAN.from_hparams(source="speechbrain/tts-hifigan-ljspeech", savedir=tmpdir_vocoder)
>>> mel_specs = torch.rand(2, 80,298)
>>> waveforms = hifi_gan.decode_batch(mel_specs)

>>> # You can use the vocoder coupled with a TTS system
>>> # Intialize TTS (tacotron2)
>>> tmpdir_tts = getfixture('tmpdir') / "tts"
>>> tacotron2 = Tacotron2.from_hparams(source="speechbrain/tts-tacotron2-ljspeech", savedir=tmpdir_tts)
>>> # Running the TTS
>>> mel_output, mel_length, alignment = tacotron2.encode_text("Mary had a little lamb")
>>> # Running Vocoder (spectrogram-to-waveform)
>>> waveforms = hifi_gan.decode_batch(mel_output)

HPARAMS_NEEDED = ['generator']

decode_batch(spectrogram)[source]

Computes waveforms from a batch of mel-spectrograms

Parameters: spectrogram (torch.tensor) – Batch of mel-spectrograms [batch, mels, time]
Returns: waveforms – Batch of mel-waveforms [batch, 1, time]
Return type: torch.tensor

decode_spectrogram(spectrogram)[source]

Computes waveforms from a single mel-spectrogram

Parameters

spectrogram (torch.tensor) – mel-spectrogram [mels, time]

Returns

waveform (torch.tensor) – waveform [1, time]
audio can be saved by
>>> waveform = torch.rand(1, 666666)
>>> sample_rate = 22050
>>> torchaudio.save(str(getfixture(‘tmpdir’) / “test.wav”), waveform, sample_rate)

forward(spectrogram)[source]: Decodes the input spectrograms

training: bool