Kartoffel-Orpheus 3B (German, natural) — GGUF (ggml-quantised)
A German fine-tune of
canopylabs/orpheus-3b-0.1-ft, trained primarily on
natural human speech recordings from German speakers. Drop-in checkpoint swap on the same Orpheus runtime — same Llama-3.2-3B-Instruct talker arch, same SNAC 24 kHz codec, same
<custom_token_N> super-frame protocol, just different LM weights and a German speaker roster. The synthetic-data sibling lives at
cstr/kartoffel-orpheus-3b-german-synthetic-GGUF.
| Speakers | Names |
|---|
| Male | Jakob, Anton, Julian, Jan, Alexander, Emil, Ben, Elias, Felix, Jonas, Noah, Maximilian |
| Female | Sophie, Marie, Mia, Maria, Sophia, Lina, Lea |
Pair this with the SNAC codec at
cstr/snac-24khz-GGUF — the talker outputs codec tokens but doesn't render audio without it.
Files
| File | Quant | Size | Notes |
|---|
kartoffel-orpheus-de-natural-f16.gguf | F16 | ~6.2 GB | Reference quality |
kartoffel-orpheus-de-natural-q8_0.gguf | Q8_0 | ~3.4 GB | Recommended |
kartoffel-orpheus-de-natural-q4_k.gguf | Q4_K | ~1.8 GB | Smallest; smoke-tested but Q8_0 preferred |
Like the upstream Orpheus, sub-Q8 quants tend to break the SNAC super-frame slot pattern on rare prompts, so we ship Q4_K with a recommendation to use Q8_0 by default.
Quick start
1# 1. Build CrispASR
2git clone https://github.com/CrispStrobe/CrispASR
3cd CrispASR
4cmake -B build -DCMAKE_BUILD_TYPE=Release
5cmake --build build -j --target crispasr-lib
6
7# 2. Pull the talker + the SNAC codec
8huggingface-cli download cstr/kartoffel-orpheus-3b-german-natural-GGUF \
9 kartoffel-orpheus-de-natural-q8_0.gguf --local-dir .
10huggingface-cli download cstr/snac-24khz-GGUF snac-24khz.gguf --local-dir .
11
12# 3. Synthesise — German prompt with a German speaker
13./build/bin/crispasr --backend kartoffel-orpheus \
14 -m kartoffel-orpheus-de-natural-q8_0.gguf \
15 --codec-model snac-24khz.gguf \
16 --voice Julian \
17 --temperature 0.6 \
18 --tts "Hallo, ich heiße Julian und das ist ein Kartoffel-Orpheus Test." \
19 --tts-output hallo.wav
For auto-download simply pass -m auto:
1./build/bin/crispasr --backend kartoffel-orpheus-de-natural -m auto \
2 --voice Sophie --temperature 0.6 \
3 --tts "Auto-download holt beide Dateien." \
4 --tts-output out.wav
Quality verification
ASR roundtrip via
cstr/parakeet-tdt-0.6b-v3-GGUF on Q8_0, voice
Julian:
| Synthesised text | parakeet-v3 -l de output |
|---|
"Hallo, ich heiße Julian und das ist ein Kartoffel-Orpheus Test." | "Hallo, ich heiße Julian und das ist ein Kartoffel-Orpheus-Test." (verbatim, only minor hyphenation drift) |
Validation script:
1crispasr --backend kartoffel-orpheus-de-natural \
2 -m kartoffel-orpheus-de-natural-q8_0.gguf \
3 --codec-model snac-24khz.gguf --voice Julian --temperature 0.6 \
4 --tts "Hallo, ich heiße Julian und das ist ein Kartoffel-Orpheus Test." \
5 --tts-output kartoffel_test.wav
6crispasr --backend parakeet -m parakeet-tdt-0.6b-v3-q4_k.gguf -l de \
7 -f kartoffel_test.wav --no-prints
8# → Hallo, ich heiße Julian und das ist ein Kartoffel-Orpheus-Test.
Architecture
Identical to Orpheus 3B — see
cstr/orpheus-3b-0.1-ft-GGUF for the full architecture writeup. The CrispASR
orpheus runtime is checkpoint-agnostic; this GGUF is loaded by the same
orpheus_init_from_file path with no source-code changes.
Prompt format (verbatim from the upstream): the LM sees [audio_start=128259, BOS=128000, ...tokenize("{name}: {text}")..., eot_id=128009, audio_eot=128260, audio_eom=128261, audio_end=128257]. The runtime handles this for you; just pass --voice <Name>.
Stop policy
Stop on audio_end=128257 or on >4 consecutive non-codec tokens. The audio_pre_end=128009 and audio_end_b=128261 tokens are not termination signals — they overlap with Llama-3 specials in the prompt and get filtered silently by the runtime.
Conversion
1python models/convert-orpheus-to-gguf.py \
2 --input SebastianBodza/Kartoffel_Orpheus-3B_german_natural-v0.1 \
3 --output kartoffel-orpheus-de-natural-f16.gguf \
4 --speakers Jakob,Anton,Julian,Jan,Alexander,Emil,Ben,Elias,Felix,Jonas,Noah,Maximilian,Sophie,Marie,Mia,Maria,Sophia,Lina,Lea \
5 --variant fixed_speaker
6
7build/bin/crispasr-quantize kartoffel-orpheus-de-natural-f16.gguf \
8 kartoffel-orpheus-de-natural-q8_0.gguf q8_0
9build/bin/crispasr-quantize kartoffel-orpheus-de-natural-f16.gguf \
10 kartoffel-orpheus-de-natural-q4_k.gguf q4_k
The --variant fixed_speaker flag bakes the German speaker roster into orpheus.fixed_speakers so the runtime's --voice <Name> lookup works without an external mapping.
Attribution
- Talker base:
SebastianBodza/Kartoffel_Orpheus-3B_german_natural-v0.1. German fine-tune by Sebastian Bodza on natural German speech.
- Upstream Orpheus base:
canopylabs/orpheus-3b-0.1-ft.
- Llama base:
meta-llama/Llama-3.2-3B-Instruct — Llama-3.2 community license.
- SNAC codec:
hubertsiuzdak/snac_24khz (MIT) — see cstr/snac-24khz-GGUF.
- GGUF conversion + ggml runtime:
CrispStrobe/CrispASR — see src/orpheus.cpp, src/orpheus_snac.cpp, models/convert-orpheus-to-gguf.py.
License
Llama-3.2 community license (inherited from the talker base). Includes the Acceptable Use Policy and the "Built with Llama" attribution requirement. Commercial use is permitted under the community license terms.
The SNAC codec is MIT and ships separately under
cstr/snac-24khz-GGUF.
Provenance and EU AI Act Art. 53 note
- Upstream model: SebastianBodza/Kartoffel_Orpheus-3B_german_natural-v0.1 — published by
SebastianBodza.
- Upstream licence:
llama3.2. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
- What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
- Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository. No training-content summary was found on the upstream model card at the time of writing; that documentation gap is upstream's and is not filled here.
- Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.
