martinturuta/safi-kinyarwanda-conversations

Safi Diction Kinyarwanda Conversational Speech Dataset

This dataset contains 1 hour of Kinyarwanda conversational speech collected using Safi's collection engine.

The recordings contain multiple speakers responding to survey questions. The original recordings were processed using speaker diarization to identify speaker turns. Consecutive turns from the same speaker were consolidated and split into speaker-specific audio clips of up to 15 seconds. These clips were then transcribed and paired with speaker, timestamp, demographic, and audio-quality metadata.

Both the original recordings and processed speaker-level clips are included. This preserves the full source conversations while also providing speaker-separated data suitable for speech and language modeling tasks.

This is just a sample for development or testing - contact hq@safidata.com for the full dataset or custom datasets, or visit https://www.safidata.com/.

Dataset Creation Process

The dataset was produced using the following pipeline:

1. Audio collection — Multi-speaker survey responses were collected remotely using Safi's collection engine.

2. Speaker diarization — Each recording was processed using pyannote/speaker-diarization-community-1 to identify speaker turns.

3. Speaker-turn consolidation — Consecutive turns attributed to the same speaker were combined where possible.

4. Clip generation — Consolidated speaker regions were divided into clips with a maximum duration of 15 seconds.

5. Short-turn filtering — Speaker regions shorter than 1 second were omitted from the processed dataset.

6. Transcription — The resulting diarized speaker clips were transcribed.

7. Audio-quality annotation — Each transcribed clip was assigned an audio-quality assessment.

8. Transcription filtering — Clips without a non-empty transcription were excluded from the final processed dataset.

9. Metadata enrichment — Speaker demographic information collected during the survey, including gender and age range where available, was associated with the corresponding speakers.

10. Sentence-aligned annotations — Transcriptions were associated with speakers and timestamps relative to the original recording.

The processing pipeline can be summarized as:

Original recording
    ↓
Speaker diarization
    ↓
Speaker turns
    ↓
Consecutive same-speaker turns consolidated
    ↓
Maximum 15-second speaker clips
    ↓
Transcription
    ↓
Audio-quality annotation
    ↓
Filtering
    ↓
Speaker + timestamp + demographic metadata

Dataset Structure

The repository contains both the original recordings and processed diarized clips:

sentence_aligned_dataset/
├── original_audio/
│   ├── sample_0001_q5_1.wav
│   ├── sample_0002_q5_2.wav
│   └── ...
│
├── audio_clips/
│   ├── sample_0001_q5_1_SPEAKER_00_000000588_000008384.wav
│   ├── sample_0001_q5_1_SPEAKER_01_000008620_000010021.wav
│   └── ...
│
├── annotations/
│   ├── sample_0001_q5_1.json
│   ├── sample_0002_q5_2.json
│   └── ...
│
├── manifest.jsonl
└── dataset_summary.json

Original Audio

original_audio/ contains the original multi-speaker survey recordings before diarization and segmentation.

These files preserve the complete conversational context and can be used to reproduce, inspect, or evaluate the diarization and segmentation process.

Diarized Audio Clips

audio_clips/ contains speaker-specific clips generated from the original recordings.

Each filename identifies:

• the original recording
• the diarized speaker
• the start timestamp
• the end timestamp

For example:

sample_0001_q5_1_SPEAKER_00_000000588_000008384.wav

represents a segment from:

sample_0001_q5_1.wav

assigned to:

SPEAKER_00

The numeric timestamps encoded in the filename represent milliseconds from the start of the original recording.

In this example:

000000588 → 0.588 seconds
000008384 → 8.384 seconds

Annotation Files

annotations/ contains one JSON file for each original recording.

Each annotation contains:

• recording metadata
• source audio information
• speaker mapping information
• coverage information
• chronological speech segments
• sentence-level timing
• transcriptions
• audio-quality annotations
• participant identifiers
• speaker-level demographic information
• full speaker-level transcriptions

The annotation type is:

speaker_diarized_sentence_aligned

All timestamps use:

seconds_from_start_of_original_recording

as their reference.

Example Annotation

A simplified annotation has the following structure:

{
  "schema_version": "1.2",
  "recording_id": "sample_0001_q5_1",
  "respondent_id": "1",

  "source_audio": {
    "filename": "sample_0001_q5_1.wav",
    "relative_path": "original_audio/sample_0001_q5_1.wav"
  },

  "annotation_type": "speaker_diarized_sentence_aligned",
  "timing_reference": "seconds_from_start_of_original_recording",

  "speaker_mapping": {
    "method": "first_speaker_appearance",
    "verified": false,
    "mapping": {
      "SPEAKER_00": "person_1",
      "SPEAKER_01": "person_2"
    }
  },

  "coverage": {
    "first_included_second": 0.588,
    "last_included_second": 36.481,
    "included_duration_seconds": 34.965
  },

  "segments": [
    {
      "segment_id": "sample_0001_q5_1_SPEAKER_00_000000588_000008384",
      "segment_index": 1,
      "speaker_id": "SPEAKER_00",

      "source_clip": {
        "filename": "sample_0001_q5_1_SPEAKER_00_000000588_000008384.wav",
        "relative_path": "audio_clips/sample_0001_q5_1_SPEAKER_00_000000588_000008384.wav"
      },

      "timing": {
        "start_seconds": 0.588,
        "end_seconds": 8.384,
        "duration_seconds": 7.796
      },

      "transcript": {
        "text": "Nkurikirane mucyumweru gisanzwe mukazi kawe cyangwa mu bucuruzi bwawe umunsi usanzwe uba umeze ute?",
        "sentences": [
          {
            "sentence_id": "sample_0001_q5_1_SPEAKER_00_000000588_000008384_sentence_001",
            "sentence_index": 1,
            "text": "Nkurikirane mucyumweru gisanzwe mukazi kawe cyangwa mu bucuruzi bwawe umunsi usanzwe uba umeze ute?",
            "start_seconds": 0.588,
            "end_seconds": 8.384,
            "duration_seconds": 7.796
          }
        ]
      },

      "audio_quality": {
        "raw_value": "Loud Background Noise",
        "labels": [
          "loud_background_noise"
        ]
      },

      "person_id": "person_1",
      "participant_id": "1_person_1"
    }
  ],

  "speakers": [
    {
      "speaker_id": "SPEAKER_00",
      "person_id": "person_1",
      "participant_id": "1_person_1",

      "demographics": {
        "gender": "male",
        "gender_raw": "Umugabo",
        "age_range": "26-40",
        "raw_value": "Umugabo, 26-40"
      },

      "full_text": "...",
      "segment_count": 2,

      "segment_ids": [
        "sample_0001_q5_1_SPEAKER_00_000000588_000008384",
        "sample_0001_q5_1_SPEAKER_00_000030440_000036481"
      ]
    }
  ]
}

Segment Metadata

Each item in segments corresponds to one processed diarized audio clip.

Field	Description
segment_id	Unique identifier derived from the clip filename.
segment_index	Chronological position of the segment within the recording.
speaker_id	Speaker label assigned during diarization.
person_id	Person identifier within the survey response.
participant_id	Unique participant identifier combining respondent and person IDs.
source_clip.filename	Filename of the diarized audio clip.
source_clip.relative_path	Relative repository path to the clip.
timing.start_seconds	Segment start time in the original recording.
timing.end_seconds	Segment end time in the original recording.
timing.duration_seconds	Duration of the segment.
transcript.text	Transcription associated with the clip.
transcript.sentences	Sentence-level transcription and timing information.
audio_quality.raw_value	Original audio-quality annotation.
audio_quality.labels	Normalized machine-readable audio-quality labels.

Audio Quality

Audio quality is annotated at the individual segment level.

Each segment contains:

"audio_quality": {
  "raw_value": "Loud Background Noise",
  "labels": [
    "loud_background_noise"
  ]
}

The raw_value field preserves the original quality annotation.

The labels field provides a normalized machine-readable representation.

Example values include:

clear_audio
loud_background_noise
background_noise
low_volume
clipping
distortion
overlapping_speech
inaudible
truncated

The exact labels available depend on the annotations observed in the dataset.

Audio-quality metadata can be used to filter clips or evaluate model performance under different recording conditions.

For example, users may choose to train only on:

clear_audio

or separately evaluate recognition accuracy on:

loud_background_noise

segments.

Transcription and Sentence Alignment

Each segment contains a transcript object.

For example:

{
  "transcript": {
    "text": "Hmmm ndabashimiye",
    "sentences": [
      {
        "sentence_id": "sample_0001_q5_1_SPEAKER_01_000008620_000010021_sentence_001",
        "sentence_index": 1,
        "text": "Hmmm ndabashimiye",
        "start_seconds": 8.62,
        "end_seconds": 10.021,
        "duration_seconds": 1.401
      }
    ]
  }
}

This dataset provides segment/sentence-level timing rather than word-level alignment.

The sentence timestamps correspond to positions in the original recording.

Word-level timestamps were not inferred or artificially generated.

Speaker Metadata

The speakers array contains recording-level metadata for each diarized speaker.

Each speaker entry contains:

• speaker_id
• person_id
• participant_id
• demographic information
• full combined transcription
• number of included segments
• IDs of all segments associated with that speaker

Example:

{
  "speaker_id": "SPEAKER_01",
  "person_id": "person_2",
  "participant_id": "1_person_2",

  "demographics": {
    "gender": "female",
    "gender_raw": "Umugore",
    "age_range": "26-40",
    "raw_value": "Umugore, 26-40"
  },

  "full_text": "Hmmm ndabashimiye ...",
  "segment_count": 3,

  "segment_ids": [
    "sample_0001_q5_1_SPEAKER_01_000008620_000010021",
    "sample_0001_q5_1_SPEAKER_01_000010713_000025713",
    "sample_0001_q5_1_SPEAKER_01_000025713_000030440"
  ]
}

Participant Identifiers

Three related identifiers are used in the dataset.

respondent_id

Identifies the survey response or recording group.

For example:

respondent_id = 1

Multiple speakers may share the same respondent_id because they participated in the same survey response.

person_id

Identifies a person within a survey response.

For example:

person_1
person_2

participant_id

Provides a participant-level identifier by combining the respondent and person identifiers.

For example:

1_person_1
1_person_2

This makes it possible to distinguish individual speakers while retaining their relationship to the original survey response.

Demographic Metadata

Where available, speaker demographic information includes:

Field	Description
gender	Normalized English gender label.
gender_raw	Original Kinyarwanda survey value.
age_range	Reported age bucket.
raw_value	Original combined demographic response.

For example:

{
  "gender": "male",
  "gender_raw": "Umugabo",
  "age_range": "26-40",
  "raw_value": "Umugabo, 26-40"
}

and:

{
  "gender": "female",
  "gender_raw": "Umugore",
  "age_range": "26-40",
  "raw_value": "Umugore, 26-40"
}

Speaker Mapping

Speaker labels generated by the diarization model are arbitrary.

For example:

SPEAKER_00
SPEAKER_01

do not inherently identify specific participants.

The annotation therefore contains an explicit speaker_mapping object:

{
  "speaker_mapping": {
    "method": "first_speaker_appearance",
    "verified": false,
    "mapping": {
      "SPEAKER_00": "person_1",
      "SPEAKER_01": "person_2"
    }
  }
}

The mapping method used in this dataset is:

first_speaker_appearance

This means diarized speakers were associated with survey participant positions based on their order of first appearance in the recording.

Importantly:

verified = false

indicates that this mapping was not manually verified for every recording.

Users performing participant-specific or demographic analyses should take this limitation into account.

Coverage

Each annotation includes a coverage object describing how much of the original recording is represented in the processed dataset.

For example:

{
  "coverage": {
    "first_included_second": 0.588,
    "last_included_second": 36.481,
    "included_duration_seconds": 34.965
  }
}

These fields indicate:

Field	Description
first_included_second	Timestamp of the first included transcribed segment.
last_included_second	End timestamp of the final included segment.
included_duration_seconds	Total duration of all included segments.

Because very short regions and clips without usable transcriptions may be removed, the included duration may be shorter than the full original recording.

Speaker Diarization

Speaker diarization was performed using:

pyannote/speaker-diarization-community-1

The raw diarization output was post-processed before generating the final clips.

Consecutive turns assigned to the same speaker were merged when separated by no more than approximately:

2 seconds

A speaker change always ended the current merged region.

Short raw turns were retained during this merging stage so that they could become part of a valid longer same-speaker region.

After merging, regions shorter than:

1 second

were removed.

Long speaker regions were then divided into clips with a maximum duration of:

15 seconds

A clip never crosses a detected speaker change.

Clip Generation

The diarized clip naming convention is:

{source_stem}_{speaker}_{start_ms}_{end_ms}.wav

For example:

sample_0001_q5_1_SPEAKER_01_000010713_000025713.wav

means:

Source recording:
sample_0001_q5_1.wav

Speaker:
SPEAKER_01

Start:
10.713 seconds

End:
25.713 seconds

Duration:
15.000 seconds

This naming convention makes every clip directly traceable to its position in the original recording.

Transcription Filtering

Only diarized clips containing a non-empty transcription are included in the final processed dataset.

The audio_clips/ directory therefore represents the subset of diarized clips for which usable transcription data is available.

The complete source recordings remain available in original_audio/.

Cleaning

The final dataset applies several processing and cleaning rules:

• consecutive nearby turns from the same speaker were consolidated
• speaker changes were preserved as hard boundaries
• merged regions shorter than 1 second were removed
• speaker regions longer than 15 seconds were split
• clips without non-empty transcriptions were excluded
• audio-quality information was retained for included clips
• original recordings were preserved alongside processed data

The diarized clips should therefore be considered derived, processed data rather than raw recordings.

Manifest

manifest.jsonl provides a recording-level index of the dataset.

Each line corresponds to one original recording.

Example:

{
  "recording_id": "sample_0001_q5_1",
  "respondent_id": "1",
  "source_audio_filename": "sample_0001_q5_1.wav",
  "source_audio_path": "original_audio/sample_0001_q5_1.wav",
  "annotation_path": "annotations/sample_0001_q5_1.json",
  "segment_count": 5,
  "speaker_count": 2,
  "included_duration_seconds": 34.965,
  "speaker_mapping_method": "first_speaker_appearance",
  "speaker_mapping_verified": false
}

The manifest can be used as the primary entry point for loading the dataset programmatically.

Dataset Summary

dataset_summary.json contains aggregate information about the processed dataset.

This includes values such as:

• number of original recordings
• number of diarized clips
• number of annotated segments
• number of speaker labels
• total included audio duration
• annotation schema version

The summary reflects the final filtered dataset rather than the raw diarization output.

Format

The dataset contains two levels of audio.

Original recordings

Complete multi-speaker recordings:

original_audio/

Processed speaker clips

Speaker-separated segments:

audio_clips/

Annotations are stored as JSON with one annotation file per original recording.

A JSONL manifest provides a recording-level index.

All segment and sentence timestamps are expressed in seconds relative to the beginning of the original recording.

Intended Uses

The dataset structure supports tasks including:

• Kinyarwanda automatic speech recognition
• Kinyarwanda speech transcription
• speaker-aware ASR
• conversational speech recognition
• speaker diarization research
• low-resource speech modeling
• multilingual speech modeling
• speech-quality robustness evaluation
• evaluation under noisy recording conditions
• demographic analysis of speech-model performance
• development of speech-processing pipelines for conversational survey data

Users interested in full conversational context can use the original recordings.

Users interested primarily in transcription or speaker-separated speech can use the diarized clips and corresponding annotations.

Limitations

Speaker diarization is model-generated and may contain speaker-attribution or boundary errors.

The mapping between diarization labels such as SPEAKER_00 and survey participant identifiers such as person_1 uses first-speaker appearance and is marked as unverified.

Audio quality varies between recordings and segments. Quality annotations are provided so users can identify or filter clips based on recording conditions.

The dataset provides segment/sentence-level timestamps but does not provide word-level alignment.

Transcriptions are associated with processed diarized segments. Clips without usable transcriptions were excluded from the processed subset, so the processed clips do not necessarily represent every second of speech present in the original recordings.