CSM-1B — GGUF (ggml-quantised)

GGUF / ggml conversion of sesame/csm-1b (Conversational Speech Model) for use with CrispStrobe/CrispASR.

CSM-1B is a TTS model that generates speech from text using a two-stage transformer architecture:

• Backbone (Llama-3.2 1B, 16 layers): generates first-codebook Mimi tokens autoregressively
• Depth decoder (Llama-3.2 100M, 4 layers): fills remaining 31 codebooks per frame
• Mimi codec (Kyutai, 8-layer transformer + SEANet): converts 32-codebook RVQ tokens to 24 kHz PCM

Released under Apache 2.0 license.

Files

Quick start

# 1. Build CrispASR
git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j --target crispasr-cli

# 2. Download model
huggingface-cli download cstr/csm-1b-GGUF csm-1b-q8_0.gguf --local-dir .

# 3. Synthesize
./build/bin/crispasr --backend csm -m csm-1b-q8_0.gguf \
    --tts "Hello, how are you today?" \
    --tts-output hello.wav --seed 42

Or with auto-download:

./build/bin/crispasr -m csm --auto-download \
    --tts "The quick brown fox jumps over the lazy dog." \
    --tts-output fox.wav

Parameters

Architecture details

• Text tokenizer: Llama-3.2 BPE (128,256 tokens)
• Audio codec: Mimi (32 codebooks, 2048 entries each, 12.5 Hz frame rate)
• Backbone: 16-layer Llama with GQA (32 heads, 8 KV heads), SwiGLU, RMSNorm, RoPE theta=500,000
• Depth decoder: 4-layer Llama (8 heads, 2 KV heads), position-specific codebook heads
• Sample rate: 24,000 Hz

Conversion

python models/convert-csm-to-gguf.py \
    --input sesame/csm-1b \
    --output csm-1b-f16.gguf

# Quantize
./build/bin/crispasr-quantize csm-1b-f16.gguf csm-1b-q8_0.gguf q8_0
./build/bin/crispasr-quantize csm-1b-f16.gguf csm-1b-q4_k.gguf q4_k

Acknowledgements

• SesameAILabs/csm — original model and inference code
• Kyutai/mimi — audio codec

Provenance and EU AI Act Art. 53 note

• Upstream model: sesame/csm-1b — published by sesame.
• Upstream licence: apache-2.0. 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.
• 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.