rlabz/qsuperposition_mwanamke

Swahili Culture Conversational Dataset (Stereo)

Overview

This dataset contains conversational Swahili audio, restructured into diarized, speaker-separated stereo chunks suitable for fine-tuning speech-to-speech dialogue models — specifically prepared for fine-tuning Moshika (the female-voice Moshi variant) via kyutai-labs/moshi-finetune.

The processing pipeline was adapted from the approach used in rlabz/sample_sw_culture.

Source Data

• Origin: r-labs/kencorpus_sw_culture — a Swahili, Culture-genre audio dataset built from Kencorpus/KenCorpus_audio (CC-BY-4.0)
• Source file: a single ~1 hour 5 minute (1:05:40) source recording (audio/35.wav), 48kHz, stereo, 16-bit PCM
• Language: Swahili (Kiswahili)
• License: CC-BY-4.0 (inherited from the source dataset — attribution to the KenCorpus authors required for any use of this dataset)

Processing Pipeline

1. Speaker diarization — The source recording was processed with pyannote/speaker-diarization-community-1 to detect speaker turns and boundaries.
2. Conversational chunking — The full recording was split into conversational chunks based on speaker turn-taking.
3. Stereo channel construction — For each chunk, a two-channel (stereo) clip was built from the mono source: left channel contains only one speaker's segments (silence elsewhere), right channel contains only the other 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).

⚠️ Important: Stereo Audio Decoding

The audio column contains genuine 2-channel stereo audio. The Hugging Face datasets library's built-in Audio feature automatically downmixes multi-channel audio to mono on decode — accessing example["audio"]["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/qsuperposition_sw_culture", split="train")
ds = ds.cast_column("audio", Audio(decode=False))

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

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

Dataset Fields

Field	Description
audio	Stereo audio clip: left = one speaker's segments, right = the other speaker's segments (decode manually, see above)
duration	Chunk duration in seconds

Intended Use

Prepared for fine-tuning speech-to-speech / conversational speech models — specifically Moshika (female-voice Moshi) via kyutai-labs/moshi-finetune — on Swahili dialogue.

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.
• This is a small sample derived from a single ~1 hour source recording, intended as an initial test set for the fine-tuning pipeline rather than a full training corpus.

License & Attribution

This dataset is derived from r-labs/kencorpus_sw_culture (itself built from Kencorpus/KenCorpus_audio) and is distributed under CC-BY-4.0. Please attribute the original KenCorpus authors when using this dataset.