Downloads · 30 days
0
becausereasons/yue2-blgr-bulgarian-voices
yue2-blgr-bulgarian-voices is a text-to-audio model from becausereasons. Use it for the text-to-audio task on the model card, and read the license before you ship it in a product. The card lists the license as cc-by-nc-4.0.
Three artist-style LoRAs that give YuE2-3B a vocal texture it does not have: the Bulgarian women's folk choir (open-throat singing, close dissonant seconds, drones, ornamented leads), and, in two of the three files, T…
Downloads · 30 days
0
Access
Public
Updated Nov 3, 2026
Repo size
565 MB
Likes
7
Public
Click a slice to open those files.
.safetensors532 MB · 94%
From the Hugging Face model README
Three artist-style LoRAs that give YuE2-3B a vocal texture it does not have: the Bulgarian women's folk choir (open-throat singing, close dissonant seconds, drones, ornamented leads), and, in two of the three files, Tuvan throat singing (kargyraa growl, overtone whistle). They were built to be fused: the choir sits happily on trap, jungle, house and ambient prompts, and every plan comes back in the same dark modal centre.
Each file patches both halves of YuE2 in one go: the autoregressive planner (writes the score, decides the arrangement and the vocal lines) and the flow-matching decoder (the sound). Trigger word for all three: blgr.
| LoRA | File | Character | Start here | |
|---|---|---|---|---|
| ⛰️ | BLGR Rhodope | blgr_rhodope.safetensors | The choir. Widest, most reliable, and the one for fusion: a cappella, trap, jungle, tribal house. No throat singing in it. | clip 1.0 / model 1.0, cfg 1.0 |
| 🐎 | BLGR Steppe | blgr_steppe.safetensors | Choir and throat voice trained half and half, earlier checkpoint. The pick for duets where both voices appear. | clip 1.0 / model 1.5, wild recipe |
| 🔥 | BLGR Kargyraa | blgr_kargyraa.safetensors | Same run, final checkpoint. The throat voice is strongest here; push the decoder. | clip 1.0 / model 1.5–2.0, cfg 1.0 |
ComfyUI workflows: download both: a core ComfyUI version (built-in nodes only, no custom nodes) and an FS_Audio Suite version. Each has a setup note and the recommended settings for every LoRA in this YuE2 series.
All demos: seed 7, 32 steps dpm_2 / sgm_uniform, no post-processing, original Bulgarian folk-style lyric (the duet and the throat solo use a version whose bridge is written in Tuvan).
BLGR Rhodope — a cappella choir, prompt prompts/a_cappella_choir.txt. The sanity check: no instruments asked for, 90 % of vocal bars sung (96 BPM, D♯ minor, 3:26):
<audio controls src="https://huggingface.co/becausereasons/yue2-blgr-bulgarian-voices/resolve/main/demos/blgr_rhodope__a_cappella.mp3"></audio>
BLGR Rhodope — trap choir, prompt prompts/trap_choir.txt: choir hooks over a half-time trap beat (2/4, 70 BPM, D♯ minor, 2:32):
<audio controls src="https://huggingface.co/becausereasons/yue2-blgr-bulgarian-voices/resolve/main/demos/blgr_rhodope__trap_choir.mp3"></audio>
BLGR Rhodope — jungle choir, prompt prompts/jungle_choir.txt: ornamented female lead over chopped breaks. The only plan in the whole series that left D♯ minor, and the densest vocal (175 BPM, F minor, 2:06, 95 % of bars sung):
<audio controls src="https://huggingface.co/becausereasons/yue2-blgr-bulgarian-voices/resolve/main/demos/blgr_rhodope__jungle_choir.mp3"></audio>
BLGR Rhodope — trap soul, prompt prompts/trap_soul_duet.txt, wild recipe: slow trap drums, electric piano, silky choir harmonies (66 BPM, D♯ minor, 3:02):
<audio controls src="https://huggingface.co/becausereasons/yue2-blgr-bulgarian-voices/resolve/main/demos/blgr_rhodope__trap_soul.mp3"></audio>
BLGR Steppe — choir and throat duet, prompt prompts/choir_throat_duet.txt, wild recipe, decoder 1.5: choir on the verses and chorus, the throat voice named for the Tuvan-language section (90 BPM, D♯ minor, 4:03):
<audio controls src="https://huggingface.co/becausereasons/yue2-blgr-bulgarian-voices/resolve/main/demos/blgr_steppe__choir_throat_duet.mp3"></audio>
BLGR Kargyraa — electronic ritual, prompt prompts/electronic_ritual.txt, baseline recipe, decoder 1.5: both voices over slow broken-beat drums, metallic percussion, granular choir stutters and cold modular arpeggios (90 BPM, D♯ minor, 3:12, 79 % of vocal bars sung):
<audio controls src="https://huggingface.co/becausereasons/yue2-blgr-bulgarian-voices/resolve/main/demos/blgr_kargyraa__electronic_ritual.mp3"></audio>
BLGR Kargyraa — throat solo, prompt prompts/throat_solo.txt, baseline recipe, decoder 2.0: the throat voice alone, a cappella (96 BPM, D♯ minor, 4:00):
<audio controls src="https://huggingface.co/becausereasons/yue2-blgr-bulgarian-voices/resolve/main/demos/blgr_kargyraa__throat_solo.mp3"></audio>
BLGR Kargyraa — throat drone, prompt prompts/throat_drone.txt, decoder 1.5: a short study. The lyric is nothing but long Tuvan vowels under a [Spoken] tag, the prompt asks for one sustained low pitch and no melody (96 BPM, D♯ minor, 0:48):
<audio controls src="https://huggingface.co/becausereasons/yue2-blgr-bulgarian-voices/resolve/main/demos/blgr_kargyraa__throat_drone.mp3"></audio>
Ready-made workflows: download them here. The core ComfyUI version needs no custom nodes (core LoRA loader + YuE2 nodes, same settings as below); the FS_Audio Suite version is the chain below.
The LoRAs are in the fused-key layout that Comfy's YuE2 implementation uses (text_encoders.* for the planner, diffusion_model.* for the decoder). They were trained with, and load through, the FS_Audio Suite node pack.
yue2_3b_bf16.safetensors from Comfy-Org/YuE2 in models/checkpoints/.blgr_*.safetensors into models/loras/.🧩 FS_Audio Lora Loader ──loras──▶ 🎤 FS_Audio Model Loader ──pipe──▶ 🎵 FS_Audio Sampler ──▶ 💿 FS_Audio Output
lora_name = blgr_rhodope.safetensors yue2_checkpoint = yue2_3b_bf16 style = <prompt, starts with "blgr,">
strength_clip = 1.0 (planner) melody_transcriber = none lyrics = <tagged lyric blocks, Cyrillic>
strength_model = 1.0 (decoder) score_mode = full
Sampler settings used for every demo:
| widget | value |
|---|---|
| steps / sampler / scheduler | 32 / dpm_2 / sgm_uniform |
| score_mode | full |
| song_length_cap | 300 (plans run 2–4.5 minutes; 240 for the throat solo) |
| repetition_penalty | 1.2 |
Recipes:
| Baseline | Wild | Throat | |
|---|---|---|---|
| strength_clip (planner) | 1.0 | 1.0 | 1.0 |
| strength_model (decoder) | 1.0 | 1.0–1.5 | 1.5–2.0 |
| Weirdness (cfg) | 1.0 | 1.4 | 1.0 |
| score_temperature | 0.7 | 0.9 | 0.7 |
| music_temperature | 1.0 | 1.2 | 1.0 |
Headless, the same graph as an API prompt:
prompt = {
"1": {"class_type": "FSAudioLoraLoader", "inputs": {"lora_name": "blgr_rhodope.safetensors", "strength_model": 1.0, "strength_clip": 1.0}},
"2": {"class_type": "FSAudioModelLoader", "inputs": {"yue2_checkpoint": "yue2_3b_bf16.safetensors", "melody_transcriber": "none", "loras": ["1", 0]}},
"3": {"class_type": "FSAudioSampler", "inputs": {"pipe": ["2", 0], "style": STYLE, "lyrics": LYRICS, "seed": 7,
"song_length_cap": 300, "score_mode": "full", "steps": 32, "sampler": "dpm_2", "scheduler": "sgm_uniform",
"Weirdness (cfg)": 1.0, "score_temperature": 0.7, "music_temperature": 1.0, "repetition_penalty": 1.2}},
"4": {"class_type": "FSAudioOutput", "inputs": {"song": ["3", 0], "score": ["3", 1], "info": ["3", 2], "filename_prefix": "blgr"}},
}
Start with the trigger, then one descriptive sentence: language → genre → vocal → instruments → mood → production → BPM. Language is Bulgarian for the choir and Bulgarian and Tuvan (or Tuvan) when you want the throat voice.
blgr, Bulgarian, traditional Balkan folk choral, a cappella female choir, penetrating open-throat polyphonic vocals, sustained drone voices, close dissonant harmonies, haunting and mystical, natural hall reverb
Ready-made prompts in prompts/:
| file | what it gets you |
|---|---|
a_cappella_choir.txt | the pure choir (demo) |
trap_choir.txt | half-time trap with choir hooks (demo) |
jungle_choir.txt | 170 BPM breaks with an ornamented lead (demo) |
trap_soul_duet.txt | slow trap soul with choir harmonies; use the wild recipe (demo) |
tribal_house.txt | four-on-the-floor with tupan and darbuka polyrhythms, call-and-response choir; wild recipe. A favourite that is not in the demo set |
choir_throat_duet.txt | choir verses, throat voice on the Tuvan section (demo, Steppe) |
throat_solo.txt | the throat voice alone (demo, Kargyraa) |
throat_drone.txt | one sustained low pitch, no melody; pair it with a lyric of long vowels (Өөөө ааааа хөөөө) under a [Spoken] tag. Renders come out short (demo, Kargyraa) |
throat_harsh.txt | the same, described as a harsh gravelly growl with a rattling sub-bass fundamental and whistle overtones. Writes a different, more song-like plan and holds the low register better; decoder 1.5 |
electronic_ritual.txt | experimental electronic ritual: broken-beat drums, granular choir stutters, modular arpeggios, both voices; decoder 1.5 (demo, Kargyraa) |
deep_house_choir_throat.txt | deep house with both voices; wild recipe, decoder 1.5. The one prompt whose plan kept the Tuvan section as a separate bridge |
What fuses and what does not. Rhythmic, percussive electronic vocabulary renders well: four-on-the-floor, breakbeats, hi-hat rolls, tupan, darbuka, rim clicks, dub stabs, pads. Prompts that lean on synthetic bass design (wobble bass, acid sequences, harsh noise sweeps) are the ones most likely to come back as mush; the trap and jungle demos work because the drums carry them. Add "sung throughout" to the vocal description, otherwise electronic prompts drift toward long instrumental stretches.
Three things together, on Steppe or Kargyraa:
[Bridge] of four to six Tuvan lines inside an otherwise Bulgarian lyric is what the duet demo uses.Be realistic about it: this is a throat colour learned from sixteen minutes of audio, not a specialist model. It is recognisably there and gets closer with the three steps above; it is not as dark and rasping as a field recording.
Tagged blocks in Cyrillic, short lines, folk-song repetition. [Intro], [Verse 1], [Chorus], [Verse 2], [Chorus], [Bridge], [Outro]. Vocable lines (Ой, ой, ой, Дай, дай, дай) work as intros and bridges, the way the tradition uses them. Shape that made the demos:
[Intro]
two vocable lines
[Verse 1]
four short lines, six to nine syllables
[Chorus]
four lines, first and third identical
[Verse 2]
four short lines
[Chorus]
(repeat, written out)
[Bridge]
two vocable lines, or four to six lines in Tuvan for the throat singer
[Outro]
the chorus image, two lines
Trained on a single RTX 5090 with the FS_Audio Suite Artist Trainer on Comfy's own YuE2 weights. One run trains the planner LoRA and the decoder LoRA together; each file also carries full vae2llm / llm2vae projection diffs. Rhodope is one run; Steppe and Kargyraa are two checkpoints of a second run.
| Rhodope | Steppe | Kargyraa | |
|---|---|---|---|
| dataset | 54 items, 156 min: Bulgarian women's choir folk repertoire (mostly a cappella, a few with folk instruments), plus 7 excerpts from a choir / Tuvan ensemble collaboration | 35 items, 79 min: 17 slow choir chants, 17 throat-singing excerpts of ~55 s, 1 self-rendered anchor | same |
| checkpoint | step 350 (final) | step 250 | step 350 (final) |
| planner / decoder steps | 350 / 600 | 350 / 600 | 350 / 600 |
| artist (planner) loss | 5.472 → 4.635 | 5.355 → 4.452 | 5.355 → 4.424 |
| decoder loss at checkpoint | 1.069 (monotone from 1.145) | 1.044 (the minimum) | 1.049 |
| wall time | 21 min | 19 min |
Hyper-parameters: rank 64 planner / rank 32 decoder, LR 3e-5 planner / 4e-5 decoder / 2e-5 I/O projections, artist fraction 0.5 vs regularizer pack, KL 0.1, batch 2 songs, 8192 max tokens, 750-frame windows, EMA 0.99, score-first 0, no transcription.
The first run had the throat-singing excerpts at about a tenth of the audio. Result: the choir was excellent and the throat voice was simply absent. Asked for a throat solo, the planner wrote zero sung bars. A second run with cleaner data at the same ratio gave the planner the mode (96 % of bars sung on the same prompt) but the decoder still never produced the timbre.
The fix was not more data and not duplicated files. The trainer picks a training item uniformly at random and then a random 30-second window inside it, so a voice's share of training is its share of items, not of minutes. Cutting the four long throat excerpts into sixteen distinct pieces of about 55 seconds, and cutting the choir side to its seventeen most typical slow chants, made the set half and half by item count (one fifth by minutes) without repeating a second of audio. That run is Steppe / Kargyraa, and it is the first where the throat voice came through by ear.
Other things that mattered:
[Instrumental] lyric file rather than a guessed transcription.blgr_rhodope.safetensors 177 MB choir run, step 350 — planner + decoder LoRA (bf16 weights, fp32 projection diffs)
blgr_steppe.safetensors 177 MB choir + throat run, step 250
blgr_kargyraa.safetensors 177 MB choir + throat run, step 350
demos/ mp3 renders (192 kbps from the FLAC masters), seed 7 throughout
prompts/ style prompts
These LoRAs are trained on my own GPU and released free. If they're useful to you and you'd like to chip in for compute, there's a Ko-fi: ko-fi.com/becausereasons <3
Weights are released under CC BY-NC 4.0, inherited from the YuE2-3B base model. Non-commercial use only; attribute "BLGR LoRAs by becausereasons".