QaraamiGenAI

Where the research is now

Dhageyso — listen

4 clips generated by the project's research models, each labelled with the model and the recordings its adapter learned from. Every clip is AI output, not a recording, and none of the source recordings is served here. The models are work in progress: they have learned the timbre and melodic shape of the tradition, and they still add noise and drift from the scale at times. The numbers under each clip are from the project's own scorer, unedited.

Selection: the 4 clips below are the top 4 of 16 clips the model produced for held-out prompts, ranked by an automatic score (pitch conformity plus trackable melody) — a curated selection, not a random draw. The source recordings are not distributed.

  1. Solo oud, 110 BPM, rooted on G

    AI output

    Prompt: qaraami, Somali traditional song, led by the oud (kaban), solo oud, moderate at 110 BPM, pentatonic melody rooted on G, studio recording, clean, solo instrumental, no vocals

    PCS
    0.974
    Voiced
    77 %
    Band SNR
    18.9 dB
    Length
    30 s

    AI model output — generated, not a recording · ACE-Step 1.5 xl-sft (4 B) · clean-oud LoRA (v3_ace_xlsft, final), trained on isolated oud stems from one performer’s own recordings in a private collection, 5.7 hours; no archival or internet audio

  2. Solo oud, 110 BPM, rooted on C

    AI output

    Prompt: qaraami, Somali traditional song, led by the oud (kaban), solo oud, moderate at 110 BPM, pentatonic melody rooted on C, studio recording, clean, solo instrumental, no vocals

    PCS
    0.901
    Voiced
    84 %
    Band SNR
    14.5 dB
    Length
    30 s

    AI model output — generated, not a recording · ACE-Step 1.5 xl-sft (4 B) · clean-oud LoRA (v3_ace_xlsft, final), trained on isolated oud stems from one performer’s own recordings in a private collection, 5.7 hours; no archival or internet audio

  3. Solo oud, 107 BPM, rooted on E

    AI output

    Prompt: qaraami, Somali traditional song, led by the oud (kaban), solo oud, moderate at 107 BPM, pentatonic melody rooted on E, studio recording, clean, solo instrumental, no vocals

    PCS
    0.983
    Voiced
    68 %
    Band SNR
    13.8 dB
    Length
    30 s

    AI model output — generated, not a recording · ACE-Step 1.5 xl-sft (4 B) · clean-oud LoRA (v3_ace_xlsft, final), trained on isolated oud stems from one performer’s own recordings in a private collection, 5.7 hours; no archival or internet audio

  4. Solo oud, lively at 150 BPM, rooted on G

    AI output

    Prompt: qaraami, Somali traditional song, led by the oud (kaban), solo oud, lively at 150 BPM, pentatonic melody rooted on G, tuned 41 cents flat of A440, studio recording, clean, solo instrumental, no vocals

    PCS
    0.983
    Voiced
    67 %
    Band SNR
    14.6 dB
    Length
    30 s

    AI model output — generated, not a recording · ACE-Step 1.5 xl-sft (4 B) · clean-oud LoRA (v3_ace_xlsft, final), trained on isolated oud stems from one performer’s own recordings in a private collection, 5.7 hours; no archival or internet audio

Recordings

No recording is served on this site yet. The archive's source recordings are held for research: the Harvard Loeb Music Library collection is research-only and never appears here, and the privately shared oud (kaban) collection waits for its written consent record before any of it is published. Rights-clear recordings will be listed in this section, each with its source stated, as they are cleared.

What the numbers mean

PCS (pentatonic conformity score): the share of voiced frames whose pitch sits within ±50 cents of a degree of the five-note scale fitted to the clip. Voiced: the share of frames with a confident pitch; a low value means long unpitched stretches, not silence. Band SNR: the clip's signal-to-noise ratio in the melody band; the adapters still add hiss, and this is the number that shows it. All three come from the same scorer used in the evaluation, not from listening.