Default model-manager downloads use the published GGUF package when available; the original source/conversion instructions below remain valid for manual use.
inflect_v2 provides native GGML inference for the English
Inflect Micro v2 and
Inflect Nano v2 TTS models.
Both variants use the same runtime and produce 24 kHz mono audio. Inference is offline; the published packages use FP32 weights.
The runtime does not use ONNX Runtime. The default download is the published GGUF package. The deprecated converter can still rebuild local safetensors from the official ONNX exports for manual testing.
Inflect v2 uses the shared eSpeak-ng phonemizer, including shared synchronization with other model frontends.
With AUDIOCPP_STATIC_ESPEAK=ON, the build includes eSpeak code and stages its
data beside the CLI/server; no separate installation is needed.
Otherwise install eSpeak-ng and its English voice data first. On Debian or Ubuntu:
sudo apt install espeak-ng libespeak-ng1On macOS:
brew install espeak-ngThen install the default GGUF package:
python tools/model_manager_v2.py install inflect_micro_v2_orig --models-root modelsThe deprecated source-conversion workflow remains available through
tools/model_manager_deprecated.py for users who want to rebuild from the
upstream ONNX assets themselves.
audiocpp_cli --task tts --family inflect_v2 \
--model models/Inflect-Micro-v2-GGUF/inflect-micro-v2-orig.gguf --backend cpu \
--text "Hello from Inflect Micro version two." \
--seed 0 --out inflect.wavUse a converted Nano checkpoint to run Nano. CUDA uses the same model layout and request surface. CUDA sessions keep duration alignment on the native CPU GGML path so small TF32 differences cannot change the monotone expansion; the four flows and waveform decoder execute on CUDA.
If eSpeak-ng is installed outside the dynamic loader's search path, provide the library and data paths explicitly:
audiocpp_cli --task tts --family inflect_v2 \
--model models/Inflect-Micro-v2-GGUF/inflect-micro-v2-orig.gguf --backend cpu \
--session-option inflect_v2.espeak_library_path=/path/to/libespeak-ng.so \
--session-option inflect_v2.espeak_data_path=/path/to/espeak-ng-data \
--text "A configured eSpeak installation." --out inflect.wavOn Windows, espeak-ng.dll and libespeak-ng.dll are searched automatically.
If eSpeak-ng is not installed system-wide, espeakng-loader can provide the
external DLL and data directory:
$espeakPaths = uv run --with espeakng-loader==0.2.4 python -c `
"import espeakng_loader; print(espeakng_loader.get_library_path()); print(espeakng_loader.get_data_path())"
$espeakLibrary = $espeakPaths[0]
$espeakData = $espeakPaths[1]
.\audiocpp_cli.exe --task tts --family inflect_v2 `
--model ../../models/Inflect-Micro-v2 --backend cpu `
--session-option "inflect_v2.espeak_library_path=$espeakLibrary" `
--session-option "inflect_v2.espeak_data_path=$espeakData" `
--text "Hello from Inflect Micro version two." `
--seed 0 --out inflect.wavPython is used only to locate the external files. Synthesis remains entirely inside the native audio.cpp process.
| Field | Value |
|---|---|
| Family | inflect_v2 |
| Tasks | tts |
| Modes | offline |
| Languages | en |
| Default package | inflect_micro_v2_orig |
| Input | Text; voice conditioning as shown in the examples |
| Output | Generated speech audio |
“Model/session default” means no explicit override. The runtime uses the checkpoint or session setting, or leaves the optional value unset as described.
| Option | Meaning |
|---|---|
--model |
Model file or directory shown in the examples. |
--backend |
Select an available backend supported by this model. |
--threads |
CPU thread count. |
--log |
Enable runtime logging. |
--text |
Input text; see the examples for the expected prompt format. |
--out |
Output WAV file. |
| Option | Type / values | Default | Description |
|---|---|---|---|
speed |
float |
1.0 |
Speech speed multiplier mapped directly to Inflect speed; default 1.0. Declared range: minimum 0.5, maximum 2.0. |
variation |
float |
0.667 |
Latent Gaussian variation; default 0.667. Declared range: minimum 0.0, maximum 1.0. |
seed |
int |
0 |
Non-negative latent noise seed; default 0. Declared range: minimum 0. |
text_chunk_mode |
word_budget |
word_budget |
Long-form text chunking mode. |
text_chunk_size |
int |
280 |
Maximum Unicode codepoints per long-form text chunk; default 280. Declared range: minimum 1. |
| Option | Type / values | Default | Description |
|---|---|---|---|
inflect_v2.espeak_library_path |
path |
Model/session default | Optional explicit path to the eSpeak-ng shared library. |
inflect_v2.espeak_data_path |
path |
Model/session default | Optional explicit path to the directory containing espeak-ng-data. |
inflect_v2.exec_plan |
string |
{} |
JSON object of component execution plans. Component: decoder. Each plan accepts impl, backend, artifacts, and params. impl is default/ggml/rknn. GGML keeps the session backend. RKNN requires artifacts.main and ENGINE_ENABLE_RKNN; omit backend. Neither implementation takes params. |
Long-form text is split near punctuation. Adjacent chunks receive 5 ms edge fades and punctuation-dependent pauses. A request is rejected if its expanded latent would exceed 4000 frames. The model has one fixed voice and exposes no language, speaker, cloning, or streaming controls.
The inflect_v2 model specification supports standalone GGUF packaging. Pack
the FP32 model.safetensors together with config.json:
audiocpp_gguf \
--input weights=models/Inflect-Nano-v2/model.safetensors \
--root models/Inflect-Nano-v2 \
--output models/Inflect-Nano-v2-FP32/model.gguf \
--type orig --family inflect_v2 \
--model-spec model_specs/inflect_v2.jsoneSpeak-ng remains an external runtime dependency and is never embedded in the GGUF.
FP16 and quantized packages are intentionally unsupported until separate parity and listening validation is available.
- English, one fixed voice, and offline TTS only.
- eSpeak-ng and its data remain external runtime dependencies.
- Only FP32 packages are supported.
- Windows CPU/CUDA and Linux CPU under WSL2 are runtime-validated. Native Linux CUDA, Vulkan, and Metal are not practically validated for this release.
- CUDA deliberately performs duration alignment on CPU. CPU and CUDA are deterministic within a backend but are not bit-identical to each other.
- Windows WDDM did not expose per-process VRAM through NVML, so no CUDA peak VRAM figure is claimed.