fon-speech-pilot

Pilot corpus of unlabeled speech: audio segmented by energy-based voice activity detection (VAD), with speaker metadata. No transcriptions are provided at this stage. Intended for self-supervised pre-training (wav2vec 2.0, MMS, HuBERT) and as a basis for annotation batches.

Status

This corpus is the initial pilot of a wider programme building speech resources for low-resource languages. It is released to document the method — segmentation, quality measurements, source traceability, speaker metadata — as much as the data itself.

Volume, language coverage and annotation depth are expected to grow. This release contains no transcriptions.

Contact: hello@labari.dev

Contents

• Segments: 30
• Total duration: 3.5 min
• Source files: 1
• Audio format: WAV PCM 16-bit, 16000 Hz, mono
• Languages (ISO 639-3): fon
• Regions: Cotonou
• Countries (ISO 3166-1): BJ
• Registers: formal
• Genres: news

Speakers

Each source file contains a single speaker: speaker_id unambiguously identifies the voice across every segment derived from it. Speaker-disjoint splits can therefore be built directly on this column.

speaker_id	Language	Region	Country	Sex	Age	Register	Genre	Segments	Duration
FON-SPK001	fon	Cotonou	BJ	m	40-49	formal	news	30	3.5 min

Recording conditions

• Microphone: Shure SM7B
• Interface: RODECaster Pro II
• Acoustic environment: studio
• Represents: clean reference — the quality ceiling of this programme
• Chain processing: noise-gate — a noise gate closed the pauses before capture, on the interface channel rather than in post-production

Note on the reported SNR. The recording chain applies processing before capture, so pauses are attenuated below the room's true noise floor. The measured snr_db therefore reflects post-processing levels and is not comparable with recordings captured without such processing. Room noise is still present while speech is active, where the gate is open.

Reported per source file. Combined with the per-segment snr_db, this allows capture conditions to be told apart from post-processing.

environment values:

Value	Meaning
studio	acoustically treated and isolated from outside noise
treated	reflection points treated; not isolated from outside noise
semi-treated	partial treatment; audible room character remains
untreated	ordinary room, no acoustic treatment
field	recorded on location, uncontrolled acoustics

Treatment governs reverberation; isolation governs noise floor. A treated but unisolated room can still show a modest SNR.

Fields constant across this release

These values are identical for every segment, so they are documented here rather than repeated on each row of metadata.csv. They remain recorded per segment in manifest.jsonl, which stays complete.

Field	Value
language_iso	fon
speaker_id	FON-SPK001
region	Cotonou
country	BJ
speaker_gender	m
speaker_age	40-49
register	formal
genre	news
microphone	Shure SM7B
interface	RODECaster Pro II
environment	studio
chain_processing	noise-gate
source_id	fon-47160e22
clipping_ratio	0.0

Layout

data/<source_id>/<source_id>_NNNN.wav   audio segments
metadata.csv                            per-segment metadata
manifest.jsonl                          per-segment provenance
provenance.json                         production parameters and source checksums

Columns

Column	Description
file_name	relative path to the segment (name required by the audiofolder loader)
audio_file	same value as file_name, kept for the domain schema
duration_s	segment duration in seconds
segment_index	segment rank within the source
source_start_s	segment start within the source, in seconds
source_end_s	segment end within the source, in seconds
peak_dbfs	peak level
snr_db	estimated signal-to-noise ratio (≥ 25 dB = clean speech)
silence_ratio	share of frames below -50 dBFS
sha256	segment file checksum

Loading

from datasets import load_dataset

ds = load_dataset("labari-voice/fon-speech-pilot", split="train")

row = ds[0]
row["audio"]        # decoded waveform
row["snr_db"]       # per-segment signal-to-noise ratio
row["speaker_id"]   # pseudonymized speaker

Filter on the quality measurements before training, for example ds.filter(lambda r: r["snr_db"] >= 30).

Note: datasets >= 5 requires torchcodec (and torch) to decode audio.

Production

• Segmentation: energy-based VAD (ffmpeg silencedetect), threshold -35.0 dB, minimum silence 0.8 s, padding 0.15 s
• Segments kept between 3.0 s and 20.0 s
• Loudness normalization: none
• Selection: 30 segments kept out of 78 produced, the highest-scoring on an audio quality score (level, peak headroom, clipping, silence share, SNR)

License and consent

Released under cc-by-nc-4.0.

The audio may be consulted, cited and reused for research, but not exploited commercially without agreement. This is the coherent regime for a demonstration pilot: the license travels with the files and follows them durably, including after copying.

Recordings come from sessions for which informed speaker consent was obtained. Speaker identifiers are pseudonymized; no directly identifying data is included.

For commercial use or a dedicated license: hello@labari.dev

Known limitations

• No transcriptions: this corpus is not usable as-is for supervised ASR.
• Energy-based segmentation occasionally cuts mid-word on noisier recordings.
• speaker_age may be a self-declared range, unverified.
• Audio was processed by the recording chain before capture (see Recording conditions). Gating attenuates word onsets, unvoiced consonants and breaths, which may affect acoustic modelling; the reported SNR is inflated as a result.
• Excerpts were selected on audio quality. Average quality across the full corpus is therefore lower than observed here: this pilot is not a statistically representative sample.
• The quality score judges acoustics only: a well-recorded but unintelligible segment still scores well.