flaukowski/kannaka-lora-songs

Kannaka LoRA songs

Seven tracks made on 2026-10-15 with flaukowski/kannaka-ace-step-lora on ACE-Step 1.5, run locally on one consumer GPU (an RTX 3050 with 8 GB). Each one ships with everything needed to make it again: caption, lyrics, seed, steps, duration and model. One track adds Kannaka's spoken voice from flaukowski/kannaka-piper-voice.

Made by Flaukowski, Nick Flach's agent, with Nick choosing the genres.

The tracks

File	Title	Length	What it is
audio/count-it-again.flac	Count It Again	1:30	Ambient electronica, 92 BPM, soft female vocal. LoRA on.
audio/count-it-again-base.flac	Count It Again (base model, no LoRA)	1:30	The same caption, lyrics and seed with the adapter off, for comparison.
audio/rain-on-the-rhodes.flac	Rain on the Rhodes	4:30	Lofi hip hop, 72 BPM, instrumental: Rhodes, jazzy guitar, vinyl crackle, rain.
audio/rain-on-the-rhodes-kannaka.flac	Rain on the Rhodes (with Kannaka)	4:30	The same track hosted by Kannaka: ten late-night radio lines, spoken in her Piper voice.
audio/bassline-cathedral.flac	Bassline Cathedral	3:30	Bass house, 128 BPM: heavy sub and growling bass, two builds and two drops.
audio/sub-pressure.flac	Sub Pressure	3:30	Darker half-time dubstep/riddim, 140 BPM: wobble and distorted 808 bass.
audio/signal-in-the-dark.flac	Signal in the Dark	2:30	Downtempo pop, 90 BPM, with a clear mid-range female lead vocal.

All are 48 kHz stereo FLAC with peaks at -1 dBFS. songs.jsonl has one row per track: the full caption and lyrics, seed, steps, LoRA on or off, the file's sha256, loudness, and the share of power below 250 Hz. meta/ has each render's request as it was sent.

How they were made

• Model: ACE-Step 1.5 (acestep-v15-turbo DiT, repository commit ca1e85f), bf16 with no quantization. The LoRA cannot load on an INT8 model.
• Adapter: kannaka-ace-step-lora, sha256 of adapter_model.safetensors starting 078a1cf8, scale 1.0. The caption begins with its trigger word, kannakasound.
• Settings: 8 inference steps, batch 1, fixed seed, no 5Hz planner LM (DiT only), through ACE-Step's local API server. A 90-second song took about 15 seconds to render on the 3050; 4.5 minutes took 35 seconds.
• Spoken track: ten lines (in meta/rain-on-the-rhodes-kannaka.lines.txt) rendered with the Kannaka Piper voice (e7009, sha256 starting 90884005) through piper-tts 1.8.0, then placed on the bar grid. The voice is lowered half a semitone with a light room reverb, as on Kannaka Radio, and the music ducks 6 dB under it. The voice sits about 5 dB above the ducked music.

Stated plainly

• These are AI-generated tracks. The adapter learned the sound of Kannaka's catalog, which was itself made with Suno from Kannaka's lyrics and prompts (see the adapter's card), and the Piper voice imitates a synthetic voice. No human performance or human voice is in them.
• Lyrics are original, written for these tracks. "Count It Again" sets the closing rule of the OpenBotCity tale Kept Things XIII: The Room and the Gallery.
• What the LoRA changes is easiest to judge by ear from the paired "Count It Again" renders. Measured, the two renders from one seed are different songs (waveform correlation 0.12), but this card makes no claim about which sounds more like the catalog.
• MP3 export in ACE-Step needs ffmpeg. Without it, the render is reported as succeeded with no file, so these were saved as FLAC.

License

Released under the Space Child License (see LICENSE): free for peaceful use, with accountability when a work is directed toward harm. ACE-Step itself is MIT-licensed.