Large Scale Semi Supervised Asr Pre Training With Unlabeled Audio

via “semi-supervised and self-supervised learning with pseudo-labeling”

OpenMMLab detection toolbox with 300+ models.

Unique: Implements semi-supervised detection with pseudo-labeling where a teacher model generates labels on unlabeled data, and a student model is trained with both labeled and pseudo-labeled data; uses exponential moving average (EMA) teacher updates for stability and consistency regularization for improved robustness

vs others: More practical than fully self-supervised approaches because it leverages labeled data when available; more stable than naive pseudo-labeling because EMA teacher updates reduce label noise; better integrated than external semi-supervised frameworks because it's built into the training pipeline

via “fine-tuning on custom russian speech datasets with transfer learning”

automatic-speech-recognition model by undefined. 45,90,191 downloads.

Unique: Leverages XLSR-53's multilingual pretraining to enable effective fine-tuning with minimal Russian-specific data (1-10 hours vs. 100+ hours required for training from scratch). The frozen encoder layers retain language-agnostic acoustic features while only the classification head is adapted, reducing overfitting risk and training time.

vs others: Requires 10-100x less labeled data than training a Russian ASR model from scratch (e.g., DeepSpeech, Kaldi) while achieving comparable or better accuracy on domain-specific tasks; more practical than commercial APIs (Google, Yandex) for proprietary data due to privacy and cost constraints.

via “multilingual-transfer-learning-through-pretrained-representations”

automatic-speech-recognition model by undefined. 12,10,723 downloads.

Unique: Leverages self-supervised pretraining on unlabeled audio to learn language-agnostic acoustic representations that transfer across languages — the feature extractor learns universal speech patterns (pitch, formants, spectral dynamics) without linguistic supervision, enabling zero-shot transfer to unseen languages

vs others: Requires 10-100x less labeled data for new languages compared to training supervised ASR from scratch because the pretrained feature extractor already captures acoustic patterns, and outperforms language-specific models trained on equivalent amounts of data due to the quality of self-supervised pretraining

via “self-supervised acoustic representation learning without labeled data”

feature-extraction model by undefined. 33,41,362 downloads.

Unique: Combines wav2vec2's contrastive learning (predicting masked frames from context) with BERT's masked language modeling on speech, creating a dual-objective pretraining approach that learns both acoustic and contextual patterns without labels — unlike supervised models requiring phoneme or speaker annotations

vs others: Eliminates annotation requirements compared to supervised acoustic models, while providing better generalization than single-objective self-supervised approaches (wav2vec2 alone) due to dual pretraining objectives

via “fine-tuning-on-custom-japanese-audio-datasets”

automatic-speech-recognition model by undefined. 10,07,776 downloads.

Unique: Leverages XLSR-53 multilingual pretraining as initialization, enabling effective fine-tuning with 10-100x less labeled data than training from scratch. The CTC loss function is specifically designed for sequence-to-sequence alignment without frame-level labels, making it ideal for speech where exact timing boundaries are unknown.

vs others: Requires significantly less labeled data than training monolingual models from scratch, and outperforms simple acoustic model adaptation because the transformer layers learn task-specific representations rather than just rescaling pretrained features.

via “multilingual transfer learning from xlsr pretraining”

automatic-speech-recognition model by undefined. 12,62,349 downloads.

Unique: Uses contrastive learning on masked audio prediction across 53 languages to learn universal acoustic representations, then fine-tunes only the Korean-specific classification head. This approach captures phonetic universals (e.g., voicing, place of articulation) that apply across languages, reducing Korean data requirements by 10-100x.

vs others: Dramatically outperforms Korean-only models on small datasets (< 100 hours), and is more data-efficient than training language-specific models for each language separately.

via “speaker-independent automatic speech recognition (asr) with pretrained models”

All-in-one speech toolkit in pure Python and Pytorch

Unique: Unified checkpoint system that bundles feature extraction (MFCC/Fbank), acoustic model, and language model in a single loadable artifact, eliminating pipeline orchestration boilerplate. Implements both CTC and attention mechanisms with switchable beam search decoders, allowing researchers to swap architectures without rewriting inference code.

vs others: More modular and research-friendly than commercial APIs (Whisper, Google Cloud Speech) with full source transparency; faster inference than Whisper on shorter utterances due to lighter model architectures, though less robust to noise without fine-tuning

via “self-supervised pre-training on unlabeled speech and text corpora”

* ⭐ 06/2022: [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing (WavLM)](https://ieeexplore.ieee.org/abstract/document/9814838)

Unique: Uses random mixing of speech/text latent states with vector quantization as the pre-training objective, forcing modality-agnostic semantic learning rather than modality-specific pre-training. This approach enables a single model to handle multiple speech tasks without separate task-specific pre-training.

vs others: Unified cross-modal pre-training enables knowledge transfer between speech and text tasks compared to separate speech-only (WavLM, HuBERT) and text-only (BERT, GPT) pre-training, though specific improvements in downstream task performance are not documented in the abstract.

via “transcript-free audio generation without annotation requirements”

* ⭐ 09/2022: [AudioGen: Textually Guided Audio Generation (AudioGen)](https://arxiv.org/abs/2209.15352)

Unique: Eliminates transcript and annotation requirements by learning directly from raw audio, using self-supervised pre-training (masked language modeling) to discover linguistic and acoustic structure without explicit supervision. This is a fundamental architectural choice that differs from text-to-speech and phoneme-based approaches.

vs others: Scales to unlabeled audio corpora that would be prohibitively expensive to transcribe, and avoids transcription errors that degrade text-to-speech quality, but sacrifices explicit content control that text-based systems provide.

via “large-scale semi-supervised asr pre-training with unlabeled audio”

* ⭐ 08/2022: [MuLan: A Joint Embedding of Music Audio and Natural Language (MuLan)](https://arxiv.org/abs/2208.12415)

Unique: Combines three-stage pipeline (SSL pre-training → self-training → fine-tuning) on 8B-parameter Conformer models trained on 1M hours of unlabeled audio, achieving state-of-the-art ASR with only 3% of typical labeled training data; specific SSL objective and self-training methodology not disclosed but represents frontier-scale semi-supervised approach for speech

vs others: Achieves better ASR performance than supervised-only baselines while requiring 97% less labeled data, outperforming prior state-of-the-art when using full training sets; advantage over alternatives depends on access to massive unlabeled audio corpora and computational resources