Deployment

Inflect v2 ships as two self-contained Hugging Face model repositories with the same Python and command-line interface:

• Inflect-Micro-v2 for the stronger quality/footprint balance.
• Inflect-Nano-v2 when model size is the primary constraint.

The GitHub repository contains integration examples and documentation. The weights, frontend, runtime module, evaluation artifacts, and integrity hashes live in each Hugging Face package.

Install

git clone https://github.com/owenawsong/Inflect.git
cd Inflect
python -m pip install -r requirements.txt

The runtime requires espeak-ng. The espeakng-loader dependency supplies a portable binary for common environments; system installations are also supported by phonemizer.

Download and synthesize

python examples/download_and_speak.py \
  --model micro \
  --text "A complete local voice can fit almost anywhere." \
  --output inflect.wav

Use --model nano for the smaller package and --device cuda when CUDA is available. The first run downloads and caches the selected Hugging Face model.

Compare both models

python examples/compare_models.py \
  --text "Use the same prompt and seed for a direct comparison." \
  --output-dir comparison

This writes inflect-micro-v2.wav and inflect-nano-v2.wav with identical runtime controls.

Runtime controls

Control	Typical value	Meaning
speed	1.0	Speaking-rate multiplier. The validated range is 0.5-2.0.
variation	0.667	Stochastic delivery strength from 0.0-1.0.
seed	7	Reproduces the same stochastic generation.
device	cpu	cpu or cuda.

Long input is split at punctuation-aware boundaries, generated one chunk at a time, and joined with bounded pauses and edge fades. This limits memory use, but it is not acoustic streaming and does not provide measured time-to-first-audio.

Measured CPU profile

On the managed Hugging Face CPU Upgrade reference (8 vCPU, 32 GB RAM, four framework threads), steady-state end-to-end throughput was:

Model	Median RTF	Throughput
Inflect-Micro-v2	0.1593	6.28x real time
Inflect-Nano-v2	0.0933	10.72x real time

The protocol used 100 fixed Modern400 prompts and three passes. Pass 1 built caches and was excluded; passes 2 and 3 formed the steady-state pool. These numbers include text processing and waveform synthesis. Results will vary with hardware, thread settings, text length, and software versions.

Export formats

The validated release is PyTorch FP32. ONNX, FP16, integer-quantized, Core ML, TFLite, and GGUF packages are not currently claimed. A smaller file is not a valid release unless matched listening and intelligibility checks show that it preserves the model.