Speech recognition and audio generation

Multi-Modal Models with Hugging Face

James Chapman

Curriculum Manager, DataCamp

Speech

What are the important parts of speech waveforms?

Pitch: avg. frequency for males 85-180Hz, for females 165-255Hz
Stress: language, accent, and emotion specific
Rhythm: influenced by context and emotion and language

Social media post

Automatic speech recognition

Audio is encoded, then decoded into either audio and text.

Text and audio are both sequential
Examples: transcription, translation, text-to-speech

Automatic speech recognition

Tiny model: 39M parameters, 150MB model size
Training data: 680k hours (labeled)

from transformers import WhisperProcessor, WhisperForConditionalGeneration
processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")

¹ https://github.com/openai/whisper

Automatic speech recognition

from datasets import load_dataset, Audio
dataset = load_dataset("CSTR-Edinburgh/vctk")["train"]
dataset = dataset.cast_column("audio", Audio(sampling_rate=16_000))

sample = dataset[0]["audio"]

input_preprocessed = processor(sample["array"],
                               sampling_rate=sample["sampling_rate"], 
                               return_tensors="pt")

predicted_ids = model.generate(input_preprocessed.input_features)

transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print(transcription)

['Please cool Stella.']

Audio generation

Three components to generate audio:

Preprocessor: resampling and feature extraction
Model: feature transformation
Vocoder: a separate generative model for audio waveforms

from transformers import SpeechT5Processor, SpeechT5ForSpeechToSpeech, SpeechT5HifiGan

processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_vc")
model = SpeechT5ForSpeechToSpeech.from_pretrained("microsoft/speecht5_vc")
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")

Speech embeddings

Diagram of how a speech embedding fits in a speech generation pipeline

Generating speaker embeddings

Pre-trained encoder: audio waveform array → encoded array (typically 512 dimensions)

from speechbrain.inference.speaker import EncoderClassifier
speaker_model = EncoderClassifier.from_hparams(source="speechbrain/spkrec-xvect-voxceleb")

speaker_embeddings = speaker_model.encode_batch(torch.tensor(dataset[0]["audio"]["array"]))

speaker_embeddings = torch.nn.functional.normalize(speaker_embeddings, dim=2).unsqueeze(0)

Audio generation

inputs = processor(audio=dataset[0]["audio"], 
                   sampling_rate=dataset[0]["audio"]["sampling_rate"], 
                   return_tensors="pt")


speech = model.generate_speech(inputs["input_values"], speaker_embedding, vocoder=vocoder)

Spectrogram before and after

Let's practice!

Multi-Modal Models with Hugging Face