Upload 3 files

custom_interface.py

@@ -0,0 +1,127 @@
+import torch
+from speechbrain.pretrained import Pretrained
+
+class WhisperASR(Pretrained):
+    """A ready-to-use Whisper ASR model
+
+    The class can be used  to run only the encoder (encode()) to run the entire encoder-decoder whisper model
+    (transcribe()) to transcribe speech. The given YAML must contains the fields
+    specified in the *_NEEDED[] lists.
+
+    Example
+    -------
+    >>> from speechbrain.pretrained.interfaces import foreign_class
+    >>> tmpdir = getfixture("tmpdir")
+    >>> asr_model = foreign_class(source="hf", 
+    ... pymodule_file="custom_interface.py", 
+    ... classname="WhisperASR",
+    ... hparams_file='hparams.yaml', 
+    ... savedir=tmpdir,
+    ... )
+    >>> asr_model.transcribe_file("tests/samples/example2.wav")
+    """
+
+    HPARAMS_NEEDED = ['language']
+    MODULES_NEEDED = ["whisper", "decoder"]
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.tokenizer = self.hparams.whisper.tokenizer
+        self.tokenizer.set_prefix_tokens(self.hparams.language, "transcribe", False)
+        self.hparams.decoder.set_decoder_input_tokens(
+            self.tokenizer.prefix_tokens
+    )
+
+    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
+
+    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].
+        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.mods.whisper.forward_encoder(wavs)
+        return encoder_out
+
+    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].
+        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.mods.decoder(encoder_out, wav_lens)
+            predicted_words = self.tokenizer.batch_decode(
+                predicted_tokens, skip_special_tokens=True)
+            if  self.hparams.normalized_transcripts:
+                predicted_words = [
+                        self.tokenizer._normalize(text).split(" ")
+                        for text in predicted_words
+                    ]
+
+            
+        return predicted_words, predicted_tokens
+
+    def forward(self, wavs, wav_lens):
+        """Runs full transcription - note: no gradients through decoding"""
+        return self.transcribe_batch(wavs, wav_lens)

hparams.yaml

@@ -0,0 +1,77 @@
+# ################################
+# Model: Whisper (Encoder-Decoder) + NLL
+# Augmentation: TimeDomainSpecAugment
+# Authors: Pooneh Mousavi 2022
+# ################################
+
+
+# URL for the biggest Fairseq english whisper model.
+whisper_hub: openai/whisper-large-v2
+
+# Normalize inputs with
+# the same normalization done in the paper. Refer to Appendix C for further information.
+normalized_transcripts: True
+
+
+language: mongolian 
+
+auto_mix_prec: False
+sample_rate: 16000
+
+# These values are only used for the searchers.
+# They needs to be hardcoded and should not be changed with Whisper.
+# They are used as part of the searching process.
+# The bos token of the searcher will be timestamp_index
+# and will be concatenated with the bos, language and task tokens.
+timestamp_index: 50363
+eos_index: 50257
+bos_index: 50258
+
+# Decoding parameters
+min_decode_ratio: 0.0
+max_decode_ratio: 0.1
+test_beam_size: 8
+
+# Model parameters
+freeze_whisper: True
+freeze_encoder: True
+
+
+
+whisper: !new:speechbrain.lobes.models.huggingface_whisper.HuggingFaceWhisper
+    source: !ref <whisper_hub>
+    freeze: !ref <freeze_whisper>
+    freeze_encoder: !ref <freeze_encoder>
+    save_path: whisper_checkpoints
+    encoder_only:  False
+
+
+
+decoder: !new:speechbrain.decoders.seq2seq.S2SWhisperGreedySearch
+    model: !ref <whisper>
+    bos_index: !ref <timestamp_index>
+    eos_index: !ref <eos_index>
+    min_decode_ratio: !ref <min_decode_ratio>
+    max_decode_ratio: !ref <max_decode_ratio>
+
+# test_beam_searcher: !new:speechbrain.decoders.seq2seq.S2SWhisperBeamSearch
+#     module: [!ref <whisper>]
+#     bos_index: !ref <timestamp_index>
+#     eos_index: !ref <eos_index>
+#     min_decode_ratio: !ref <min_decode_ratio>
+#     max_decode_ratio: !ref <max_decode_ratio>
+#     beam_size: !ref <test_beam_size>
+
+
+
+
+
+modules:
+    whisper: !ref <whisper>
+    decoder:  !ref <decoder>
+ 
+
+pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
+    loadables:
+        whisper: !ref <whisper>
+

whisper.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6230026278db2e8f4cf6a49c6179bd73e1d3e3cebd0202f4615bab830029f4c5
+size 6173767281