The Best 347 Audio Classification Tools in 2025

A language identification model fine-tuned from Facebook's Massively Multilingual Speech project, supporting audio classification for 126 languages

This model is a voice command recognition model fine-tuned on the speech_commands dataset based on facebook/wav2vec2-base, achieving an accuracy of 97.59%.

A speech language identification model fine-tuned on OpenAI Whisper-medium, achieving 88.05% accuracy on the FLEURS dataset

This model is fine-tuned from Wav2Vec2-Large-Robust for speech emotion recognition, predicting values in three dimensions: arousal, dominance, and valence.

A speech language identification model based on the SpeechBrain framework and ECAPA-TDNN architecture, supporting recognition and speech embedding extraction for 107 languages.

The Audio Spectrogram Transformer (AST) is a model fine-tuned on AudioSet, which converts audio into spectrograms and applies a vision transformer for audio classification.

A general language identification model fine-tuned based on openai/whisper-small, achieving 88.6% accuracy on the evaluation dataset

Speech emotion recognition using fine-tuned wav2vec2 model, trained on IEMOCAP dataset

An audio spectrogram transformer fine-tuned on the AudioSet dataset, which converts audio into spectrograms and processes them using a vision transformer architecture, achieving excellent performance in audio classification tasks.

Gender recognition model fine-tuned on Librispeech-clean-100 dataset, achieving an F1 score of 0.9993 on the test set

English speech emotion recognition model fine-tuned based on Wav2Vec 2.0, capable of recognizing 7 different emotions

This model is a Russian speech emotion recognition model fine-tuned on the HuBERT architecture, trained on the DUSHA dataset, capable of identifying emotional states such as neutral, anger, positivity, and sadness.

MERT-v1-330M is an advanced music understanding model trained based on the MLM paradigm, with 330M parameters, supporting a 24K Hz audio sampling rate and 75 Hz feature rate, suitable for various music information retrieval tasks.

Unified automatic quality assessment model for speech, music, and sound

This is a speech language identification model based on the Wav2Vec2 architecture, capable of recognizing 256 languages, and is part of Facebook's Massively Multilingual Speech (MMS) project.

This model takes raw audio signals as input and outputs age predictions and gender probabilities (child/female/male), along with the pooled state of the last transformer layer.

A speech emotion recognition model fine-tuned on Wav2Vec 2.0, capable of identifying 8 English emotions with an accuracy of 82.23% on the RAVDESS dataset

This is a speech emotion recognition model based on the Wav2Vec2 architecture, adapted from the S3PRL project, designed to identify emotional categories in speech.

A multi-attribute speech emotion recognition baseline model based on WavLM architecture, predicting arousal, dominance, and valence dimensions

This model, fine-tuned from Wav2Vec2-Large-Robust, can predict the speaker's age and gender from raw audio.

MERT-v1-330M is an advanced music understanding model trained based on the MLM paradigm, with a parameter scale of 330M, supporting 24K Hz audio sample rate, and suitable for various music information retrieval tasks.

A pre-trained model based on the ECAPA-TDNN architecture for classifying gender from human speech

A voice content safety detection model based on WavLM base plus architecture, used to identify toxic content in voice chats

This model is a keyword spotting model based on the Hubert architecture, designed to classify speech segments into predefined keyword sets.

An audio spectrogram transformer model fine-tuned on the Speech Commands v2 dataset for audio classification tasks, achieving 98.12% accuracy.

An emotion recognition model based on Hubert-Large pre-trained model for predicting emotion categories in speech

ECAPA-TDNN architecture spoken language identification model trained on the VoxLingua107 dataset, supporting recognition of 107 languages

A synthetic speech detection model fine-tuned based on MIT/ast-finetuned-audioset-10-10-0.4593, demonstrating outstanding performance on the VoxCelebSpoof dataset

This model is an emotion recognition model based on the Hubert-Base architecture, trained on the SUPERB emotion recognition task for speech emotion classification

This project utilizes the Whisper model for speech emotion recognition, capable of classifying audio into different emotional categories such as happiness, sadness, and surprise.

This is a Persian speech emotion recognition model based on the Wav2Vec 2.0 architecture, capable of identifying six basic emotional states.

Multilingual voice toxicity detection model based on WavLM architecture, supporting 8 languages and identifying 6 types of violations

A voicemail detection model fine-tuned based on facebook/wav2vec2-xls-r-300m, specifically designed to distinguish between voicemail greetings and live human responses.

A model for classifying 20-millisecond audio frames to detect filler pauses (e.g., 'eee', 'errm', etc.)

This is a speech language identification model based on the Wav2Vec2 architecture, capable of recognizing 4017 languages, and is part of Facebook's Massively Multilingual Speech project.

A speech language identification model fine-tuned on the common_language dataset based on facebook/wav2vec2-base

This model is trained on facebook/wav2vec2-base-960h for music genre classification tasks, supporting recognition of 10 genres.

Audio classification model pre-trained on AudioSet and Librispeech, using vision transformer architecture to process audio spectrograms

This model uses the ECAPA-TDNN architecture to classify 16 accents in English speech and is trained on the CommonAccent dataset, achieving a test accuracy of 87%.

A Deepfake audio detection model fine-tuned on audio folder datasets, achieving 99.73% accuracy

Audio representation learning model based on HuBERT architecture, pre-trained on the complete AudioSet dataset

A foundational speech recognition model based on the wav2vec 2.0 architecture, pre-trained on 960 hours of English speech data