rlabz/sample_sw_culture

Swahili Culture Conversational Dataset (Stereo)

Overview

This dataset contains conversational Swahili audio samples derived from the Culture genre subset of KenCorpus_audio (CC-BY-4.0), restructured into diarized, speaker-separated stereo audio chunks suitable for fine-tuning speech-to-speech dialogue models such as Moshi / Hibiki.

Processing notebook: the full pipeline (diarization, conversational chunking, stereo construction, and upload) is available here: Google Colab notebook

Source Data

• Origin: Kencorpus/KenCorpus_audio, Culture genre subset
• License: CC-BY-4.0 (inherited from the source dataset — attribution to KenCorpus required for any use of this dataset)
• Language: Swahili (Kiswahili)

Processing Pipeline

1. Speaker diarization — Each source audio file was processed with pyannote/speaker-diarization-community-1 to detect speaker turns and boundaries.
2. Turn merging & noise filtering — Adjacent same-speaker turns were merged, and very short segments (likely diarization artifacts) were filtered out.
3. Conversational chunking — Audio was split into conversational cycles: a chunk begins when a speaker starts talking and ends when that same speaker resumes after at least one other speaker has spoken in between. This captures natural dialogue exchanges regardless of monologue length or number of interjections, without assuming strict speaker alternation.
4. Stereo channel construction — For each chunk, a two-channel (stereo) audio file was built from the mono source: left channel contains only the first speaker's segments (silence elsewhere), right channel contains only the second speaker's segments (silence elsewhere). This mirrors the channel-per-speaker structure expected by Moshi-style speech-to-speech dialogue models (similar in spirit to kyutai/DailyTalkContiguous).

Full code for every step above is in the linked Colab notebook.

⚠️ Important: Stereo Audio Decoding

This dataset's audio_stereo column contains genuine 2-channel stereo audio. However, the Hugging Face datasets library's built-in Audio feature automatically downmixes multi-channel audio to mono on decode — accessing example["audio_stereo"]["array"] the normal way will silently return a flattened mono array, even though the underlying stored file is correctly stereo.

Do not rely on the default decode. Load the column with decode=False and decode manually with soundfile to preserve both channels:

from datasets import load_dataset, Audio
import soundfile as sf
import io

ds = load_dataset("rlabz/sample_sw_culture", split="train")
ds = ds.cast_column("audio_stereo", Audio(decode=False))

def get_stereo_array(example):
    data, sr = sf.read(io.BytesIO(example["audio_stereo"]["bytes"]))
    return data, sr  # shape: (frames, 2)

If you skip this step, any downstream fine-tuning (e.g. Moshi speech-text fine-tuning) will silently train on mono audio with the two speakers merged into one channel, defeating the purpose of this dataset's structure.

Dataset Fields

Field	Description
audio_stereo	Stereo audio: left = first speaker's segments, right = second speaker's segments (decode manually, see above)
left_speaker	Diarization label of the speaker on the left channel
right_speaker	Diarization label of the speaker on the right channel
chunk_id	Index of this conversational chunk within its source file
start / end	Start/end time (seconds) of this chunk in the original source audio
duration	Chunk duration in seconds
opening_speaker	Speaker who opens this conversational cycle
turn_count	Number of merged speaker turns contained in this chunk
turns_detail	JSON list of {start, end, speaker} for every turn in the chunk (absolute timestamps, relative to the original source file)

Intended Use

Prepared for fine-tuning speech-to-speech / conversational speech models (e.g. Moshi, Hibiki) on Swahili dialogue. May also be useful for diarization evaluation or speaker-turn-aware TTS work.

Limitations

• Speaker channel separation is diarization-gated, not true acoustic source separation — during genuine overlapping speech, some bleed-through between channels can occur.
• Diarization accuracy directly affects channel assignment quality; no manual verification of speaker labels was performed at scale.

License & Attribution

This dataset is derived from Kencorpus/KenCorpus_audio and is distributed under CC-BY-4.0. Please attribute the original KenCorpus authors when using this dataset.