feat: add Soprano TTS as community model with GGUF packages

[drzsdrtfg]

Soprano TTS — Community Model

Soprano is an ultra-lightweight (~80M parameter) English-only text-to-speech model using a two-stage architecture: a Qwen3-style causal LM (17 layers, hidden 512, vocab 8192) that autoregressively emits per-frame 512-dimensional features, and a non-iterative Vocos-style decoder (ConvNeXt backbone + single ISTFT head, n_fft 2048 / hop 512) that turns those features into 32 kHz audio.

Reference: https://github.com/ekwek1/soprano
Weights: https://huggingface.co/ekwek/Soprano-1.1-80M
GGUF packages: https://huggingface.co/WalkingCat/Soprano-1.1-80M-GGUF

Files added

Path	Purpose
src/community_models/soprano_tts/ (5 .cpp)	Assets, Qwen3 LM generator, Vocos decoder, tokenizer, session
include/engine/community_models/soprano_tts/ (5 .h)	Corresponding headers
model_specs/soprano_tts.json	Schema-v1 spec, community status, GGUF + safetensors sources
docs/soprano_tts.md	User-facing documentation
docs/soprano_validation.md	Validation record with build/run commands and timing
tests/soprano_tts/soprano_warm_bench.cpp	C++ warmbench binary
tests/soprano_tts/soprano_warm_bench_cases.txt	Warmbench test cases (short/medium/long/longform)
tests/soprano_tts/soprano_python_warm_bench.py	Python reference warmbench
tools/soprano_tts/convert_soprano.py	HF checkpoint converter (folds weight-norm)
tools/soprano_tts/run_official.py	Python reference inference runner
tools/soprano_tts/compare_parity.py	Automated validation harness

Files modified

File	Change
CMakeLists.txt	+16 lines: audiocpp_add_model + add_engine_warmbench
README.md	+2 lines: community table row
docs/gguf.md	+1 line: GGUF status table row
webui/configs/models_catalog.json	+1 line: WebUI catalog entry
webui/native/dist/index.html	Rebuilt frontend with Soprano baked in

Build

# Soprano only
cmake -B build -DCMAKE_BUILD_TYPE=Release \
  -DAUDIOCPP_MODEL_SET=custom -DAUDIOCPP_MODELS=soprano_tts
cmake --build build --target audiocpp_cli --parallel

# With Vulkan
cmake -B build -DCMAKE_BUILD_TYPE=Release \
  -DAUDIOCPP_MODEL_SET=custom -DAUDIOCPP_MODELS=soprano_tts \
  -DENGINE_ENABLE_VULKAN=ON
cmake --build build --target audiocpp_cli --parallel

Quick start

# Install GGUF package
python3 tools/model_manager_v2.py install soprano_1_1_80m_q8_0

# Run inference
build/bin/audiocpp_cli --task tts --family soprano_tts \
  --model models/Soprano-1.1-80M-GGUF/soprano-1.1-80m-q8_0.gguf \
  --text "Soprano is an extremely lightweight text to speech model." \
  --out soprano.wav

Run warmbench

# Build warmbench
cmake --build build --target soprano_warm_bench --parallel

# Run (CPU)
build/bin/soprano_warm_bench --model models/soprano_pkg \
  --output-dir build/logs/warmbench/soprano_tts

# Python reference
python3 tests/soprano_tts/soprano_python_warm_bench.py \
  --model models/Soprano-1.1-80M \
  --out-dir build/logs/warmbench/soprano_tts_py

Validation

CPU performance vs official Python soprano package (transformers backend, temp=0.3, top_p=0.95):

Test	Chars	Platform	Audio (s)	Infer (s)	RTF	Speedup
short	57	Python	0.752	1.281	1.7037	—
		C++	3.136	0.740	0.2360	7.22x
medium	152	Python	2.096	1.909	0.9106	—
		C++	8.320	1.955	0.2350	3.87x
long	567	Python	7.424	5.773	0.7776	—
		C++	16.384	4.302	0.2626	2.96x

Key observations:

• C++ RTF stays below 0.27 on all tests (faster than real-time).
• Audio durations differ between Python and C++ because of different RNG implementations; both produce valid 32 kHz speech.
• Backend coverage: CPU (tested, RTF ~0.24), Vulkan (tested on RX Vega, RTF ~0.08-0.12).
• See docs/soprano_validation.md for full validation record.

Known limitations

• English-only (model limitation)
• No voice cloning
• EOS sampling unreliable at low temperature (PyTorch vs C++ RNG difference)
• Full composite build may OOM; use AUDIOCPP_MODEL_SET=custom
• Vulkan decoder output has numerical drift on AMD RX Vega (no matrix-core ops)

Code and PR assisted and written by ox alpha(GLM 5.3 flash) and Deepseek v4 flash.

[drzsdrtfg]

pls let me know If there are any issues.

[0xShug0]

@drzsdrtfg Serveral issues found:

(1) The published GGUF does not run: audiocpp_cli failed: missing tensor: decoder.head.istft.window.

(2) Converter output and spec do not match.
tools/soprano_tts/convert_soprano.py emits:

  - model.safetensors
  - decoder.safetensors

But model_specs/soprano_tts.json expects:
- combined.safetensors

(3) Documented warmbench target is missing.
gmake: *** No rule to make target 'soprano_warm_bench'. Stop.

(4) eos_bias is declared in spec/docs and used in generator, but not parsed from request options,

(5) Please follow the existing streaming implementations such as NeuTTS, Right now Soprano’s streaming path looks inconsistent with those patterns: run_mode() always returns Offline, and start_stream() parses the request options, but next_stream_event() rebuilds a default request, so streaming ignores user sampling options like seed/temperature/top_p.

[0xShug0]

@drzsdrtfg Could you isolate the changes in the framework Qwen runtime into a separate PR? It looks like a correctness fix, but it still needs validation across other models that use the same runtime to make sure there are no surprises.

[drzsdrtfg]

Something is still broken. The build doesnt work on cuda currently. Currently fixing it

[drzsdrtfg]

Oh, seems like it was another issue and you fixed it already.