cstr / bark-small-GGUF

Bark Small — GGUF

Suno Bark (MIT license) converted to GGUF for native C++ inference with CrispASR .

Model details

• Architecture : 3-stage hierarchical transformer (semantic → coarse → fine) + EnCodec decoder
• Parameters : ~300M total across 3 GPT-2 sub-models
• Output : 24 kHz mono PCM
• Languages : 13 languages with pre-trained speaker prompts
• German speakers : v2/de_speaker_0 through v2/de_speaker_9
• License : MIT

Quantization table

All variants pack the 3 sub-models (text/semantic, coarse acoustic, fine acoustic) + EnCodec decoder into a single GGUF file. No companion model needed.

Usage with CrispASR

# Auto-download and synthesize
crispasr --backend bark -m auto --tts "Hello, how are you today?" --tts-output hello.wav

# With a specific quantization
crispasr --backend bark -m bark-small-q4_k.gguf --tts "The quick brown fox" --tts-output fox.wav

# With a German speaker prompt (when supported)
crispasr --backend bark -m bark-small-q8_0.gguf --tts "Hallo Welt" --voice v2/de_speaker_3 --tts-output hallo.wav

Conversion

Produced with:

python models/convert-bark-to-gguf.py --output bark-small-f16.gguf
crispasr-quantize bark-small-f16.gguf bark-small-q8_0.gguf q8_0
crispasr-quantize bark-small-f16.gguf bark-small-q4_k.gguf q4_k

Architecture details

Stage 1 — Semantic model

• GPT-2 (12 layers, 768-d) generating semantic tokens from text
• BERT WordPiece tokenizer (119547 vocab)
• Output: up to 768 semantic tokens

Stage 2 — Coarse acoustic model

• GPT-2 (12 layers, 1024-d) converting semantic → coarse EnCodec codes
• Alternates codebook 0/1 prediction
• Output: 2 × ~384 coarse tokens

Stage 3 — Fine acoustic model

• Non-causal GPT-2 (12 layers, 1024-d)
• Fills codebooks 2-7 from codebooks 0-1
• Output: 8 codebooks × 384 timesteps

EnCodec decoder

• 8-codebook RVQ (1024 entries each)
• SEANet CNN decoder with ELU activation
• Upsample ratios [8, 5, 4, 2] → 24 kHz

Credits

• Original model: Suno AI (MIT)
• GGUF conversion + C++ runtime: CrispASR