Mumospee: A MUltiMOdal SPEEch Corpus

The Mumospee dataset supports the Meetween project’s mission of enabling inclusive, language-neutral collaboration across virtual environments. The initial release provides metadata and download scripts for a curated subset of speech audio and video sourced from publicly available datasets, optimized for processing on high-performance computing clusters.

Mumospee Overview

Mumospee is a continuously growing, comprehensive, multilingual dataset across different modalities with featuring:

• Over 561K hours speech data;
• Covering 24 EU languages and 10 non-EU languages;
• Collections of existing datasets in various speaking styles and content genres from different platforms:

_TAGS = ["CoVoST", "GigaSpeech", "PeopleSpeech", "Librispeech", "LibriTTS", "Emilia", "MOSEL"]

• A smalle version with less than 1000 rows is also available as mumospee_small for testing purpose.

The table below provides the duration statistics for each language in the dataset.

Language	Code	Count
English	en	145,997
Chinese	zh	49,922
German	de	25,145
French	fr	24,588
Italian	it	22,007
Spanish	es	21,684
Polish	pl	21,207
Dutch	nl	19,027
Czech	cs	18,705
Romanian	ro	17,906
Greek	el	17,703
Hungarian	hu	17,701
Bulgarian	bg	17,609
Portuguese	pt	17,547
Swedish	sv	16,305
Lithuanian	lt	14,400
Finnish	fi	14,200
Danish	da	13,600
Latvian	lv	13,105
Slovak	sk	12,100
Slovenian	sl	11,300
Estonian	et	10,613
Maltese	mt	9,100
Croatian	hr	8,106
Japanese	ja	1,718
Korean	ko	217
Persian	fa	110
Russian	ru	47
Welsh	cy	30
Mongolian	mn	9
Turkish	tr	8
Arabic	ar	6
Tamil	ta	4
Indonesian	id	3

Intended Uses

This dataset is designed to enable SpeechLLM and other Large Language Models to support language-neutral virtual meeting applications.

Data Sources

The initial release includes metadata and download scripts for accessing the following publicly available datasets:

• CoVoST
• GigaSpeech
• people-speech
• LibriSpeech
• LibriTTS
• Emilia
• MOSEL

Mumospee dataset structure

Mumospee is available at HuggingFace without providing all the audio data directly, but the urls or scripts to access the datasets.

In the metadata csv, each row is a sample representing the metadata of an audio, a video or a clip consisting of the following information:

• "path": the relative path of the audio file to the sample.
• "url": the link to download the parquet containing the audio, video or the clip of it.
• "type": the sample is an audio or video.
• "duration": the duration of the sample in second.
• "language": the language of the video or audio.
• "transcript": the transcript of the video or audio.
• "tag": the origin of the sample.
• "split": the sample is in split, test, or validation section in the original dataset.
• "license": the license to use this sample.

Here is an example sample:

{
    "path": "3660-172183-0000.flac",
    "url": "https://huggingface.co/datasets/meetween/mumospee_librispeech/resolve/main/librispeech-parquet/dev-other.parquet",
    "type": "audio", 
    "duration": 5.405, 
    "language": "en", 
    "transcript": "GERAINT AS HE HAD BEEN USED TO DO WHEN HE WAS AT ARTHUR'S COURT FREQUENTED TOURNAMENTS",
    "tag": "Librispeech",
    "split": "validation",
    "license": "CC-BY-4.0"
}

Example Usage

dataset= load_dataset("meetween/mumospee", trust_remote_code=True)
print(dataset)

# To get the first row of the dataset.
sample_first = dataset["train"][0]

The defaul outputs all the samples from train split. To get another splits:

# To get the dataset from test or validation split
dataset_test = load_dataset("meetween/mumospee", "test", trust_remote_code=True)
dataset_validation = load_dataset("meetween/mumospee", "validation", trust_remote_code=True)

• Filters

There are filters to select dataset samples from specific groups:

# To get the dataset of langauge "en".
dataset= load_dataset("meetween/mumospee", "test", language="en", trust_remote_code=True)

# To get the dataset from MOSEL.
dataset= load_dataset("meetween/mumospee", "train", tag="MOSEL", trust_remote_code=True)

# You can also add combination of language and tag: get English from CoVoST from test split.
dataset= load_dataset("meetween/mumospee", "test", language="en", tag="CoVoST", trust_remote_code=True)

Note: keep in mind that if a filter combination (including split) results to no dataset, you may get an value error like below:

ValueError: Instruction "train" corresponds to no data! 

Also, make sure the values are from _LANGUAGES and _TAG.

• Download audios

You can download the parquet files with the audios data by using the download_audio parameter (the default is None):

dataset= load_dataset("meetween/mumospee", "test", download_audio=True, language="en", trust_remote_code=True)

License

The metadata and download scripts are publicly available under a CC0 license. While the metadata itself is open, users must comply with the licensing terms of each underlying dataset.

---