jml2026/multilingual-accent-speech

Silencio Network: Multilingual Accent Speech Dataset (Sample)

Overview

Silencio data is valuable because it's collected in the wild from a massive, opt-in community (1.2M users across 180+ countries), giving buyers real-world accents, dialects, devices, and environments that lab or scraped datasets don't capture. Every recording is tied to explicit, traceable consent and processed with privacy-first pipelines (GDPR/CCPA compliant, anonymized, PII hashed), which reduces legal risk for enterprise buyers. On top of that, the same community lets us scale quickly into hard-to-source languages and niches, so clients get both authenticity today and a credible path to large volumes tomorrow.

This dataset is a crowdsourced multilingual–accented English and non-English speech dataset designed for model training, benchmarking, and acoustic analysis. It emphasizes accent variation, short-form scripted prompts, and spontaneous free speech. All recordings were produced by contributors using their own devices, with Whisper-generated transcripts provided for every sample.

The dataset is structured for direct use in ASR, TTS, accent-classification, diarization-adjacent analysis, speech segmentation, and embedding evaluation.

Languages and Accents

This dataset covers five language–region pairs (to find out more about other combinations please reach out to us):

• English (China): English spoken with Mandarin-influenced accent
• English (Nigeria): Nigerian-accented English
• English (United States): Native American English speakers
• German (Germany): Native German speakers
• Spanish (Mexico): Native Mexican Spanish speakers

All recordings are stored as 48 kHz WAV files.

Usage

from datasets import load_dataset

# Load a specific language-region cohort
ds = load_dataset("SilencioNetwork/multilingual-accent-speech", "spanish_mexico")

# Access different script types using splits
for sample in ds['free_speech']:
    audio = sample['audio']  # Audio data with sampling_rate and array
    transcript = sample['transcript']
    speaker_id = sample['speaker_id']

# Or access other splits
for sample in ds['keywords']:
    # Process keywords samples
    pass

for sample in ds['monologues']:
    # Process monologues samples
    pass

Speech Types

Each sample belongs to one of three categories:

• free_speech: unscripted speech on a provided topic
• keywords: short isolated prompts containing specific phrases or terms
• monologues: longer scripted passages

These values appear in the field type_of_script.

Recording Conditions

All data is crowdsourced. Contributors record themselves using their available hardware and environment; conditions therefore vary naturally across microphones, devices, and noise profiles. No studio-grade normalisation or homogenisation is applied.

Transcription

Transcriptions are machine-generated using OpenAI Whisper, preserving its segmentation structure where applicable.

Dataset Statistics

This is a sample dataset organized by language-region, with each cohort split by script type. Durations are given in hours.

Spanish (Mexico)

Split	Recordings	Speakers	Duration (hrs)
free_speech	25	5	0.27
keywords	6	2	0.05
monologues	25	7	0.45
Subtotal	56	9	0.77

English (China)

Split	Recordings	Speakers	Duration (hrs)
free_speech	25	13	0.33
keywords	25	6	0.19
monologues	25	10	0.44
Subtotal	75	18	0.96

English (Nigeria)

Split	Recordings	Speakers	Duration (hrs)
free_speech	25	23	0.32
keywords	25	23	0.16
monologues	25	21	0.53
Subtotal	75	46	1.01

English (United States)

Split	Recordings	Speakers	Duration (hrs)
free_speech	25	18	0.3
keywords	25	14	0.18
monologues	25	13	0.32
Subtotal	75	31	0.80

German (Germany)

Split	Recordings	Speakers	Duration (hrs)
free_speech	25	16	0.25
keywords	25	15	0.16
monologues	25	14	0.32
Subtotal	75	27	0.73

Overall Total

Metric	Value
Total Recordings	356
Total Speakers	161
Total Duration	4.29 hrs
Configs	5
Splits per Config	3

File Structure

Each config is organized by language-region, with splits for each script type:

spanish_mexico/
    free_speech/
        data/
            audio_389928.wav
            audio_390100.wav
            ... (25 files)
        metadata.csv
    keywords/
        data/
            audio_689765.wav
            audio_706259.wav
            ... (6 files)
        metadata.csv
    monologues/
        data/
            audio_348730.wav
            audio_348844.wav
            ... (25 files)
        metadata.csv
english_china/
    free_speech/
        data/
            audio_90018.wav
            audio_90032.wav
            ... (25 files)
        metadata.csv
    keywords/
        ... (25 files)
    monologues/
        ... (25 files)
... (3 more cohorts)

Audio files are stored separately in AudioFolder format with metadata in CSV files.

Feature Schema

All configurations share the same feature structure:

• file_name: string (relative path to audio file)
• id: integer (unique identifier)
• speaker_id: string (hashed or anonymized speaker ID)
• gender: string (speaker gender)
• ethnicity: string (speaker ethnicity)
• occupation: string (occupation or profession)
• country_code: string (ISO 3166-1 alpha-2 code)
• birth_place: string (country or region of birth)
• mother_tongue: string (native language)
• dialect: string (regional dialect)
• year_of_birth: int (birth year, YYYY)
• years_at_birth_place: int (years lived at birth place)
• languages_data: string (serialized language–proficiency data)
• os: string (recording operating system)
• device: string (recording device type)
• browser: string (browser used if web-based)
• duration: float (seconds) (audio length)
• emotions: string (emotion labels)
• language: string (primary language of the recording)
• location: string (recording location category)
• noise_sources: string (background noise labels)
• script_id: int (script template identifier)
• type_of_script: string {free_speech, keywords, monologues} (script category)
• script: string (text intended to be spoken)
• transcript: string (Whisper-generated transcription)

Licensing

Released under CC BY-NC 4.0. Commercial use is not permitted. Attribution to Silencio Network is required for any publication or derivative dataset.

Intended Use

Suitable for:

• accent-conditioned ASR training
• multilingual speech recognition
• TTS voicebank generation
• speaker embedding and similarity evaluation
• robustness benchmarking
• keyword-spotting models
• segmentation and VAD evaluation

Limitations

• Transcripts are automatically generated. Errors may be present.
• Crowdsourced device diversity introduces variable noise levels.

Contact

For questions, custom datasets, or commercial licensing inquiries, please visit our website.

Citation

@dataset{silencio_network_speech_2025,
    title        = {Silencio Network Multilingual Accent Speech Corpus},
    author       = {Silencio Network},
    year         = {2025},
    license      = {CC BY-NC 4.0}
}