Inflect-Micro-v2 Release Notes: Local TTS Under 10M Parameters

Inflect-Micro-v2 Release Notes: Local TTS Under 10M Parameters

Inflect-Micro-v2 is a local text-to-waveform speech synthesis system that achieves high-quality English voice output using only 9,356,513 deployable parameters. This release demonstrates that a complete TTS pipeline can be compressed into a 37.53 MB FP32 weight footprint while maintaining a 66.2% human preference rate over established compact baselines.

Technical Specifications and Performance

Inflect-Micro-v2 delivers 24 kHz mono output and is designed for high-efficiency local deployment on CPU or CUDA.

Model Footprint and Runtime

Inflect-Micro-v2 prioritizes quality within a strict parameter budget, while its sibling model, Inflect-Nano-v2, prioritizes the smallest possible footprint.

Metric Inflect-Micro-v2 Inflect-Nano-v2
Parameters 9,356,513 3,966,721
FP32 Weights 37.53 MB 15.97 MB
Positioning Quality-focused Footprint-focused

In CPU runtime tests using a Hugging Face CPU Upgrade instance (8 vCPU, 32 GB RAM) with four framework threads, Inflect-Micro-v2 achieved a steady-state Real-Time Factor (RTF) of 0.1593, translating to 6.28× real-time throughput.

Quality and Intelligibility Benchmarks

Evaluation was conducted using the Modern400 corpus (400 unseen English prompts). The model was measured across human preference, predicted naturalness, and ASR-based intelligibility.

  • Human Preference: Recorded a 66.2% preference rate in an anonymous community study.
  • Predicted Naturalness: Achieved a 4.395 UTMOS22 score.
  • Intelligibility: Recorded a semantic Word Error Rate (WER) of 3.99% (mean of Qwen3-ASR and Nemotron 3.5).

Architecture and Implementation

Inflect v2 is based on the VITS-family of end-to-end text-to-waveform generators. It integrates an English phoneme frontend, monotonic alignment, stochastic latent synthesis, and a residual coupling flow.

Model Components

Component Specification
Latent channels 192
Text hidden channels 96
Encoder layers / heads 3 / 2
Feed-forward channels 768
Flow coupling blocks 4
Initial decoder channels 320
Upsample rates 8, 8, 2, 2
Output 24 kHz mono waveform

Inference Controls

Users can tune the output using three primary controls:

  • Speed: Range 0.5–2.0 (Default: 1.0).
  • Variation: Range 0.0–1.0 (Default: 0.667). Lower values result in steadier speech; higher values increase latent variation.
  • Seed: Integer value used to ensure repeatable output on the same runtime stack.

Deployment and Local Execution

Inflect-Micro-v2 supports multiple deployment paths, including a native PyTorch API and a torch-free ONNX Runtime path.

PyTorch Installation

python -m pip install --upgrade huggingface_hub
hf download owensong/Inflect-Micro-v2 --local-dir Inflect-Micro-v2
cd Inflect-Micro-v2
python -m pip install -r requirements.txt

ONNX Runtime

The official FP32 export is available as Inflect-Micro-v2-ONNX. This version supports dynamic lengths and allows for provider selection (CPU/CUDA/DirectML) without requiring PyTorch.

Limitations and Scope

Inflect-Micro-v2 is an inference-first release with specific constraints:

  • Language and Voice: Supports English only with one fixed male voice. It does not support zero-shot voice cloning or multi-language adaptation.
  • Text Handling: Long passages are processed via punctuation-aware chunking rather than a single autoregressive pass, which may affect transitions.
  • Stability: Unfamiliar phrasing can lead to flatter expression or decreased stability.

Community Insights

Community feedback highlights the efficiency of the model's size relative to its output quality, though some users noted that the inflections can occasionally feel unnatural.

"The quality blow my mind for such small model! I just replaced my old onnx model with yours!"

"The inflections are weird but this doesn't sound like a robot. Not bad!"

Other observers noted that the small parameter count is encouraging for small companies or individuals to develop high-quality, compute-efficient TTS models on the Pareto frontier of quality versus runtime.

Sources