tts : implement mimi decoder

[ngxson]

llama.cpp/example/mimi

Related to #12392

This demonstrates running Kyutai's Mimi model via GGML.

TODO:

• implement decode_frame
• see how long generation goes
• test with audio codes from Sesame
• abstract the decode into a function decode(codes) that returns std::vector<float>

Quickstart

Convert model to GGUF (no need to download, the script will automatically download the safetensors file)

python examples/tts/convert_mimi_to_gguf.py

# output file: kyutai-mimi.gguf

# optionally, use q8_0 quantization for faster speed
python examples/tts/convert_mimi_to_gguf.py --outtype q8_0

Then compile, run it:

cmake --build build -j --target llama-mimi

./build/bin/llama-mimi kyutai-mimi.gguf codes.txt

# output: output.wav

# alternatively, use "dummy1" to get a "hey hello there" sample output file
./build/bin/llama-mimi kyutai-mimi.gguf dummy1

Example of code file (one code per line):

1263
1597
1596
1477
1540
1720
1433
118
1066
1968
1096
232
418
566
1653
2010

[ngxson]

Close and merge with #12648