dataset for machine learning
, , ,

A Complete Guide to best Audio Datasets

A Complete Guide to Audio Datasets

Audio Datasets

Audio Datasets. Machines are intelligent, however it is humans who take them to the incredible. Whether you need to innovate faster, provide better reporting to users, or automate and reduce fees, those solutions need expert AI teams. Equip your crew with the skills you need to get the most out of AI. courses from trusted professionals on generative AI, AWS system knowledge acquisition, voiceover engineering, LLM, NLP and much more.


 Datasets is an open source library for downloading and preparing datasets from all domains. Its minimalist API allows users to download and prepare data sets in a single line of Python code, with a set of capabilities that enable green preprocessing. The wide variety of data sets available is unparalleled, and all of the most popular machine learning data sets can be downloaded.

Not only this, but Datasets comes equipped with a couple of audio-only features that make running with audio datasets easy for researchers and professionals alike. In this blog, we’ll showcase those capabilities and show why Datasets is the ideal place to download and prepare audio datasets.


Audio Datasets

The Hub, the Face Hub that embraces, is a platform for hosting models, datasets and demos, all open source and publicly accessible. It is home to a developing series of audio datasets spanning a selection of domains, obligations and languages. Through tight integrations with Data Sets, all Hub data sets can be downloaded in one line of code.

Here, we can find additional information about the dataset, see which models specialize on the dataset, and most interestingly, listen to real audio samples. The dataset preview is presented in the center of the dataset card. It tells us the top 100 samples for each subset and divides them. In addition, the prepared audio samples are loaded so that we can listen to them in real time. If we press the play button in the main sample, we can listen to the audio and see the corresponding text.

The dataset preview is a great way to experience audio datasets before you commit to using them. You can select any data set in the Hub, scroll through the samples, and pay attention to the audio to see the various subsets and splits, evaluating whether or not it is the right data set for your needs. Once you’ve decided on a data set, it’s trivial to load the statistics so you can start using it.

Uploading an Audio Dataset, One of the key capabilities that define datasets is the ability to,download and assemble a dataset in a single line of Python code. that is made possible through the load_dataset feature. Conventionally, loading a data set includes: i) downloading the raw statistics, ii) extracting them from their compressed format, and iii) preparing individual samples and splits. With the use of load_dataset, all the heavy lifting is done under the hood.

Let’s take the example of loading the GigaSpeech dataset from Speech Colab. GigaSpeech is a remarkably recent speech recognition dataset for comparing academic speech structures and is certainly one of the many audio datasets available on Face Hub.

To load the GigaSpeech dataset, we simply take the dataset identifier in the Hub (speechcolab/gigaspeech) and specify it in the load_dataset function. GigaSpeech comes in 5 configurations of increasing duration, from xs (10 hours) to xl (10,000 hours). For the purposes of this tutorial, we will load the smallest of those configurations. The dataset identifier and preferred settings are all we need to download the dataset:

From data sets import load_dataset

gigaspeech = load_dataset(“speechcolab/gigaspeech”, “xs”)

print (gigaspeech)

Print output:


train: Dataset({

funciones: [‘segment_id’, ‘speaker’, ‘text’, ‘audio’, ‘begin_time’, ‘end_time’, ‘audio_id’, ‘title’, ‘url’, ‘source’ , ‘categoría’, ‘original_full_path’],

num_rows: 9389


validation: Dataset ({

capabilities: [‘segment_id’, ‘speaker’, ‘text’, ‘audio’, ‘begin_time’, ‘end_time’, ‘audio_id’, ‘title’, ‘url’, ‘source’, ‘category’, ‘original_full_path’ ],

num_rows: 6750


test: Dataset ({

características: [‘segment_id’, ‘speaker’, ‘text’, ‘audio’ , ‘begin_time’, ‘end_time’, ‘audio_id’, ‘title’, ‘url’, ‘source’, ‘category’, ‘original_full_path’],

num_rows: 25619

}) }


And in the same way, we have the GigaSpeech data set ready! There is no less complicated way to load an audio data set. We can see that we have the training, validation and testing splits pre-partitioned, with the corresponding statistics for each.

The gigaspeech object again via the load_dataset feature is a DatasetDict. We will treat it the same way as a normal Python dictionary. To split the teaching, we omit the key corresponding to the gigaspeech dictionary:


Print output:

Data set ({

características: [‘segment_id’, ‘speaker’, ‘text’, ‘audio’, ‘begin_time’, ‘end_time’, ‘audio_id’, ‘title’, ‘url’, ‘source’, ‘category’, ‘original_full_path’],
num_rows: 9389

This returns an element of the data set, which incorporates the information for the education division. we can cross a deeper stage and obtain the main object of division. Again, this is possible through modern Python indexing:


Print output:

{‘segment_id’: ‘YOU0000000315_S0000660’,

‘speaker’: ‘N/A’,

‘audio’: {‘path’: ‘/


00000315_S00 00660.wav’,

‘array’: array([0.0005188, 0.00085449, 0.00012207, …, 0.00125122, 0.00076294,

0.00036621], dtype= float32),

‘sampling_rate’: 16000


‘begin_time’: 2941.889892578125,

‘end_time’: 2945.070068359375,

‘audio_id’: ‘YOU0000000315’,

‘title’: ‘return to Vasselheim | crucial function: VOX MACHINA | Episode 43’,

‘url’: ‘’,

‘supply’: 2,

‘category’: 24,

‘original_full_path’: ‘audio/youtube/P0004/YOU0000000315 .opus’,


We’ll see that there are a number of features on the back of the training slice, including segment_id, speaker, text, audio, and more. For the popularity of the speech, we will deal with the text and audio columns.

Using the remove_columns method of Datasets, we will get rid of features in the dataset that are no longer needed for speech popularity:

COLUMNS_TO_KEEP = [“texto”, “audio”]

all_columns = gigaspeech[“train”].column_names

columns_to_remove = set(all_columns) – set(COLUMNS_TO_KEEP)

gigaspeech = gigaspeech.remove_columns(columns_to_remove)

Let’s check that we have correctly preserved the text and audio columns:


Print output:

{‘textual content’: “HOW CAN THEY LEAVE THEM, CAN KASH THROWS ZAHRA aside clearly quickly”,

audio’: {‘path’: ‘/home/sanchit_huggingface_co/.cache/huggingface/datasets/downloads/extracted/7f8541f130925e9b2af7d37256f2f61f9d6ff21bf4a94f7c1a3

803ec648d7d79/xs_chunks_0000 /YOU0000000315_S0000660.wav’,

‘array’: array([0.0005188, 0.00085449, 0.00012207, …, 0.00125122, 0.00076294,

0.00036621], dtype=float32),

‘sampling_rate ‘: 16000}}

top notch! We can see that we have the required 2 columns of text and audio. The text is a string with the sample transcription and the audio is a one-dimensional array of amplitude values ​​at a sampling rate of 16 KHz. It’s our loaded data set!

Easy to load, clean to process

Loading a data set with Data Sets is just half the fun. We can now use the set of equipment available to effectively preprocess our equipped data for model training or inference. In this segment, we will perform three levels of statistics preprocessing:

Audio Statistics Resampling

Preprocessing feature

Filter function

1. Resampling audio data
The load_dataset function prepares audio samples with the sample rate at which they were published. This is not always the sampling rate expected by our model. In this case, we want to resample the audio at the appropriate sampling rate.

We can set the audio inputs to our favorite sample rate using the cast_column method of Datasets. This operation does not change the audio in the area, but rather serves as an alternative indicator of the data sets to resample the audio samples on the fly as they are loaded. The following code cell will set the sampling rate to 8 kHz:

from data sets import audio

gigaspeech = gigaspeech.cast_column(“audio”, Audio(tasa de muestreo=8000))

When you reload the first audio pattern into the GigaSpeech data set, it will be resampled to the preferred sample rate:


Print output:

{‘textual content’: “WHILE THEY LEAVE, CAN KASH SEPARATES ZAHRA very quickly”,

audio’: {‘direction’: ‘/home/sanchit_huggingface_co/.cache/huggingface/datasets/downloads/extracted/7f8541f130925e9b2af7d37256f2f61f9d6ff21bf4a94f7c1a3

803ec648d7d79/xs_chunks_0000 /YOU0000000315_S0000660.wav’,

‘array’: array([ 0.00046338, 0.00034808, -0.00086153, …, 0.00099299,

0.00083484, 0.00080221], dtype=float32),

‘sampling_rate’: 8000}


We can see that the sampling rate has been reduced to 8 kHz. The values ​​in the array are also unique, since all we have now is approximately one amplitude value for each one we had before. Let’s set the data set sampling rate back to 16 kHz, the sampling rate expected by most speech recognition models:

gigaspeech = gigaspeech.cast_column(“audio”, Audio(sampling_rate=16000))


Print output:

{‘textual content’: “AS THEY LEAVE, CAN KASH SEPARATES ZAHRA very quickly”,

‘audio’: {‘curso’: ‘/home/sanchit_huggingface_co/.cache/huggingface/datasets/downloads/extracted/

3ec648d7d79/xs_chunks_0000 /YOU0000000315_S0000660.wav’,

‘array’: array([0.0005188, 0.00085449, 0.00012207, …, 0.00125122, 0.00076294,

0.00036621], dtype=float32),

‘sampling_rate ‘: 16000}


clean! cast_column provides a truthful mechanism for resampling audio data sets when necessary.

2. Preprocessing function
One of the most difficult points of working with audio data sets is preparing the data in the appropriate format for our version. Using the dataset map approach, we can write a function to pre-pattern a single pattern from the dataset and then use it on each sample without any code modification.

First, let’s load a processor object from Transformers. This processor pre-approaches the audio to the input functions and converts the target text into labels. AutoProcessor magnificence is used to load a processor from a given version checkpoint. In the example, we load the renderer from OpenAI’s Whisper medium.en checkpoint, but you can swap it for any version ID in Face Hub:

from transformers imports AutoProcesador

procesador = AutoProcessor.from_pretrained(“openai/whisper-medium.en”)

bright! Now we can write a feature that takes a single training pattern and passes it through the processor to assemble it for our model. We’ll also calculate the input duration of each audio sample, data we’ll need for the next data learning step:

def prepare_dataset(lote):

audio = lote[“audio”]

batch = processor(audio[“array”], sample_rate=audio[“sampling_rate”], text=batch[“text”])

lote[“input_length”] = len(audio[“array”]) / audio[“sampling_rate”]
devolver lote

We can practice the data preparation function in all our training examples using the data set map technique. Here, we also postponed the text and audio columns, because we have preprocessed the audio for input functions and tokenized the text into tags:

gigaspeech =, remove_columns=gigaspeech[“tren”].column_names)

3. Filtering Function
Before schooling, we would have a heuristic to filter our school records. For example, we would like to delete any audio samples longer than 30 seconds to avoid truncating them or risking out of memory errors. We will do this in batches in the same way we organized the facts for our model in the previous step.

We start by writing a function that shows which samples to keep and which to discard. This feature, is_audio_length_in_range, returns a boolean value: samples that are less than 30 seconds become true, and samples that can be longer become false.


Def is_audio_length_in_range(input_length):
devuelve input_length < MAX_DURATION_IN_SECONDS

We can follow this filtering function in all of our educational examples by using the data set elimination method, retaining all samples that are less than 30 seconds (correct) and discarding those that are longer (false):

gigaspeech[“tren”] = gigaspeech[“tren”].filtrar(is_audio_length_in_range, input_columns=[“input_length”])

And with that, we have the GigaSpeech dataset completely organized for our release! Overall, this technique required 13 strains of Python code, from loading the data set to the last filtering step.

By keeping the notebook as general as possible, we more effectively accomplish the basic steps of record instruction. However, there is no limit to the features you can observe in your audio data set. You can extend the prepare_dataset function to perform operations of much greater concern, such as data augmentation, voice activity detection, or noise reduction. With Datasets, if you can write it in a Python function, you could use it in your data set!

Streaming Mode: Silver Bullet
One of the biggest challenges facing audio data sets is their large size. GigaSpeech’s xs configuration contained only 10 hours of educational records, but accumulated over 13 GB of storage space for download and instruction.

So what happens when we need to teach about a broader division? The entire XL configuration includes 10,000 hours of educational data, requiring over 1TB of storage space. For most speech researchers, this far exceeds the specifications of a standard hard drive. May we have to shell out and buy extra garage? Or is there a way to educate on these data sets without disk space restrictions?

Data sets allow us to do precisely this. This is made possible by using streaming mode, shown graphically in parent 1. Streaming allows us to load the statistics gradually as we iterate over the data set. Instead of downloading the entire data set immediately, we load the sample of the data set using a pattern.

We iterate the data set, loading and preparing samples on the fly as they are needed. This way, we only upload the samples we use, and not the ones we don’t use! Once we are done with a sample, we keep iterating over the data set and load the next one.

This is analogous to downloading a TV show instead of streaming it. When we download a TV show, we download the entire video offline and save it to our drive. We must wait for the full video to download before we can watch it and we will need as much disk space as the size of the video file. Compare this to streaming a television show.

Here, we do not download any part of the video to disk, but instead iterate over the remote video recording and load each part in real time as needed. We don’t have to wait for the entire video to buffer before we can start watching, we’ll start once the first part of the video is ready! This is the same streaming precept we follow for uploading data sets.

father 1: Streaming mode. The data set is loaded progressively as we iterate over the data set.
Streaming mode has three main advantages over immediately downloading the entire data set:

Disk space: Samples are loaded to remember one using one as we iterate over the data set. Since statistics are not downloaded nationally, there are no disk area requirements, so you can use data sets of arbitrary size.

Download and processing time: Audio data sets are large and require a large amount of time to download and process. With streaming, loading and processing is done on the fly, meaning you can start using the data set as soon as the first sample is prepared.

24x7 offshoring transcription

Seamless experimentation: You can experiment with some samples to test that your script works without having to download the entire data set.
There is a warning about the transmission mode. When downloading a data set, both the raw data and the processed records are stored locally on disk.

If we want to reuse this data set, we can immediately load the processed logs from disk, skipping the download and processing steps. Therefore, it is best to perform the download and processing operations once, after which we can reuse the prepared logs. With streaming mode, data is not downloaded to disk.

Consequently, neither the downloaded nor the preprocessed data is cached. If we want to reuse the dataset, we need to repeat the streaming steps, loading and processing the audio documents on the fly again. Because of this, it is recommended to download data sets that you will likely use multiple times.

How will you enable streaming mode? easy! just set streaming=real when you load your data set. Relaxation can be taken care of for you:

gigaspeech = load_dataset(“speechcolab/gigaspeech”, “xs”, streaming=true)

All of the steps included up to this point in this tutorial can be applied to the streaming data set without any code changes. The simplest trade-off is that individual samples cannot be accessed using Python indexing (i.e. gigaspeech[“train”][sample_idx]). Alternatively, you should iterate over the data set, using a for loop as an example.

Streaming mode can take your research to the next level: not only are the largest data sets available to you, but you can easily compare structures across a couple of data sets in a single pass without worrying about your disk space.

Compared to evaluating a single data set, evaluating multiple data sets provides a higher metric for the generalization capabilities of a speech reputation machine (cf. End-to-End Speech Benchmark (ESB). The attached Google Colab provides an example for evaluating the Whisper version on eight English speech popularity data sets in a script using streaming mode.

A tour of the audio data sets in the Hub

This phase serves as a reference guide for the most famous voice reputation, voice translation, and audio category datasets in Face Hub. We can track everything we’ve protected for the GigaSpeech data set to any of the Hub data sets. All we need to do is pass the dataset identifier into the load_dataset function. It’s that easy!

English speaking popularity

Multilingual Voice Recognition

Voice translation

Audio category

English speech recognition
Speech reputation, or speech-to-text content, is the task of mapping speech to written textual content, where both speech and text are in the same language. We provide a summary of the most popular English speaking reputation datasets in the Hub:

Data Set Area Conversation Style Train Hours Casing Score Recommended License Usage

LibriSpeech Narrated Audiobook 960 CC-with the help of-4.0 educational reference points

non-unusual Voice 11 Wikipedia Narrated 2300 CC0-1.0 Non-local speakers

VoxPopuli ecu Parliament Oratory 540 CC0 Non-local audio system

TED-LIUM TED Talks Oratory 450 CC-through-NC-ND 3.0 Technical topics

GigaSpeech Audiobook, podcast, YouTube Narrated, spontaneous ten thousand Robustness over a couple of domains

SPGISpeech Financial meetings Public speaking, spontaneous 5000 consumer agreement absolutely formatted transcripts

income-22 Financial meetings Public speaking, spontaneous 119 range of accents

Spontaneous AMI Conferences 10 CC-via- 4.0 Noisy speech conditions

Check with Google Colab for a manual on how to compare a device across all eight English speech popularity data sets in a single script.

The subsequent data set descriptions have been largely taken from the ESB Benchmark document.

LibriSpeech ASR

LibriSpeech is a popular large-scale dataset for comparing ASR systems. It consists of approximately 1,000 hours of narrated audiobooks from the LibriVox project. LibriSpeech has been instrumental in making it easier for researchers to leverage a huge framework of pre-present transcribed speech statistics. As such, it has become one of the most famous data sets for evaluating instructional speech structures.

librispeech = load_dataset(“librispeech_asr”, “todos”)

Not unusual Voice

Not Unusual Voice is a series of publicly sourced, openly licensed speech datasets in which audio systems document textual content from Wikipedia in multiple languages. Since absolutely everyone can contribute recordings, there is great variety in both audio quality and audio systems. Audio situations are difficult, with recording artifacts, accented speech, hesitations, and the presence of foreign words. Each of the transcripts is boxed and scored.

The English subset of version 11.0 contains approximately 2,300 hours of confirmed statistics. Use of the dataset requires that you agree to the Voice common terms of use, which can be seen in Face Hub: mozilla-basis/common_voice_11_0. Once you have accepted the terms of use, you will be granted the right to access the data set. You may then need to offer an authentication token from the Hub while loading the dataset.

common_voice = load_dataset(“mozilla-basis/common_voice_11”, “es”, use_auth_token=authentic)

VoxPopuli is a large-scale multilingual speech corpus consisting of data obtained from recordings of European Parliament events from 2009-2020. Consequently, it occupies the specific domain of oratory and political discourse, largely coming from non-native audio systems. The English subset includes approximately 550 hours of labeled speech.

voxpopuli = load_dataset(“fb/voxpopuli”, “es”)

TED-LIUM is a dataset based on films of English TED speaking lectures. The talking trend is educational oratory talks. The transcribed talks cover a variety of diverse cultural, political and educational topics, resulting in a technical vocabulary. The third (today’s) edition of the data set includes approximately 450 hours of educational records. The validation and test data comes from the legacy set and is consistent with previous versions.

tedlium = load_dataset(“LIUM/tedlium”, “liberación3”)

GigaSpeech is a multi-area English speech recognition corpus curated from audiobooks, podcasts, and YouTube. It covers each narrated and spontaneous speech on a variety of topics, such as the arts, technology and sports. It consists of various fractions of schooling from 10 hours to 10,000 hours and standardized validation and verification fractions.

gigaspeech = load_dataset(“speechcolab/gigaspeech”, “xs”, use_auth_token=real)

SPGISpeech is an English speaking reputation corpus comprised of corporate earnings calls that have been manually transcribed by S&P global, Inc. The transcripts are formatted fully consistent with a professional style guide for public speaking and spontaneous speech. It includes fractions of schooling ranging from 200 hours to 5,000 hours, with canonical validation and trial divisions.

spgispeech = load_dataset(“kensho/spgispeech”, “s”, use_auth_token=proper)


Earnings-22 is a corpus of 119 hours of accumulated English-language earnings calls from international organizations. The data set was developed with the aim of aggregating a wide range of speakers and accents covering a variety of current international economic topics. There is a wide variety of speakers and accents, with speakers selected from seven exceptional language areas.

Profits-22 was generally released as a verification data set only. The Hub includes a version of the data set that has been divided into teach-validate-take a look divisions.

ganancias22 = load_dataset(“revdotcom/ganancias22”)

AMI holds up to 100 hours of meeting recordings captured using special recording sequences. The corpus contains manually annotated orthographic transcripts of the lectures aligned to the word level. The character samples in the AMI data set include very large audio documents (between 10 and 60 minutes), which can be segmented to any length possible to educate most speech reputation structures.

AMI incorporates two splits: IHM and SDM. IHM (male or female headset microphone) consists of simpler near-field speech and SDM (single distant microphone), more complex tract area speech.

ami = load_dataset(“edimburgocstr/ami”, “ihm”)

Multilingual Speech Reputation
Multilingual Speech Reputation refers to speech recognition (speech to text) for all languages ​​other than English.

LibriSpeech Multilingual
LibriSpeech Multilingual is the multilingual equivalent of the LibriSpeech ASR corpus. It comprises a massive corpus of read audiobooks drawn from the LibriVox project, making it a suitable dataset for academic studies. It contains data divided into eight over-help languages: English, German, Dutch, Spanish, French, Italian, Portuguese and Spanish.

Commonplace Voice

Common Voice is a sequence of certified, open, crowd-sourced speech datasets in which speakers record textual content from Wikipedia in multiple languages. Because everyone can contribute their recordings, there is great variation in both audio and speaker quality. Audio situations are difficult, with recording artifacts, accented speech, hesitations, and the presence of foreign words. Transcripts are in mixed case and punctuated. As of version 11, there are over 100 languages ​​available, each with low and over resources.


Untitled design 30

The voice of the people

VoxPopuli is a large-scale multilingual speech corpus that includes data from recordings of ECU Parliament events from 2009-2020. therefore, it occupies the unique domain of oratory and political discourse, largely coming from non-native speakers. Contains labeled audio transcription information for 15 EU languages.


FLEURS (Low Chance Evaluation of Popular Speech Representations Study) is a dataset for evaluating speech popularity systems in 102 languages, plus many that could be labeled as “low resource.” The information is derived from the FLoRes-101 dataset, a device translation corpus with 3001 sentence translations from English to one-zero and other languages. Native speakers are recorded narrating transcripts of sentences in their local language. Recorded audio information is combined with sentence transcriptions to generate multilingual speaking reputation in all one hundred and one languages. Educational units comprise approximately 10 hours of supervised audio transcription statistics per language.

Voice Translation
Voice translation is the task of mapping from spoken speech to written textual content, where voice and text are in different languages ​​(for example, English speech to French textual content).

CoVoST 2
CoVoST 2 is a large-scale multilingual speech translation corpus that protects translations from 21 languages ​​to English and from English to 15 languages. The dataset is created using Mozilla’s open source standard Voice database of crowd-sourced voice recordings. There are 2,900 hours of speech represented within the corpus.

FLEURS (Few-shot Gaining Knowledge Evaluation of Standard Speech Representations) is a dataset for comparing speech recognition systems in 102 languages, along with many that are classified as “low resource”. The facts are derived from the FLoRes-one zero one dataset, a machine translation corpus with 3001 sentence translations from English to one hundred and one different languages.

Local speakers are recorded narrating transcripts of sentences in their native languages. An n- way parallel corpus of speech translation information is constructed by pairing the recorded audio data with sentence transcripts for each of the one zero one languages. The educational sets contain approximately 10 hours of monitored audio transcription statistics based on the combination of source and target language.

Audio Category
The audio type is tasked with mapping a raw audio input to a category tag output. Realistic applications of the audio category consist of keyword recognition, speaker intent, and language identity.

SpeechCommands is a data set made up of one-second audio files, each containing a single spoken English word or historical noise. Words are taken from a small set of commands and spoken through several specific speakers. The data set is designed to help teach and examine small keyword recognition structures in the tool.

Multilingual Spoken Words
Multilingual Spoken Phrases is a large-scale corpus of one-second audio samples, each containing a single spoken phrase. The data set consists of 50 languages ​​and more than 340,000 keywords, with a total of 23.4 million one-second spoken examples or more than 6,000 hours of audio. Audio transcription data comes from Mozilla’s Common Voice project.

Timestamps are generated for each utterance at the word stage and used to extract spoken words from characters and their corresponding transcriptions, thereby forming a new corpus of individual spoken phrases. The intended use of the dataset is academic research and commercial applications in multilingual keyword detection and spoken term searching.

FLEURS (Low Ability Learning Evaluation of Habitual Speech Representations) is a dataset for evaluating speech reputation systems in 102 languages, including many of those categorized as “low helpful.” The records are derived from the FLoRes-101 dataset, a machine translation corpus with 3001 sentence translations from English to one zero one from different languages.

The local audio system is recorded narrating transcriptions of sentences from their native languages. The recorded audio information is combined with a tag for the language in which it is spoken. The dataset can be used as an audio class dataset for linguistic identity: structures are trained to expect the language of each utterance in the corpus.

Audio Compressor - best Ways to Reduce audio size audio quality reducer
Audio Compressor – best Ways to Reduce audio size audio quality reducer


In the concluding remarks of this blog post, we explore the adorable Face Hub and experience Dataset Preview, a powerful way to listen to audio datasets before downloading them. We load an audio data set with a line of Python code and run a series of traditional preprocessing steps to prepare it for a device to acquire knowledge of the model.

In total, this required only thirteen lines of code, relying on simple Python capabilities to perform the necessary operations. We added streaming mode, a method for loading and preparing audio data samples on the fly. We conclude by summarizing the most popular speech recognition, speech translation, and audio types datasets in the Hub.

Table of Contents