Inflect-Micro-v2: High-Fidelity Voice in 9.36M Parameters

Inflect-Micro-v2 is a compact, local text-to-waveform speech synthesis system designed to operate under a 10M parameter limit. It provides fixed-voice English TTS featuring deterministic seeds, the ability to process long-form text, and support for both CPU and CUDA inference.

A note from Owen: I independently developed and funded Inflect v2. If this version gains traction, I intend to expand the project into a broader v3, which may feature:

• Additional language support
• A wider variety of voices
• General stability enhancements

If you find this model helpful, liking it on Hugging Face significantly aids discoverability.

---

🛠 Technical Specifications

The Inflect v2 family utilizes a unified public API but is split into two distinct size tiers:

Core Metrics (Micro)

The deployable scale of the Micro model is defined as: $\text{Parameters} = 9,356,513$

---

🎧 Audio Demonstrations

The following are held-out generations (original text), not reconstructions of training data.

• Conversational: "It wasn't until later that I realized what had actually happened."
• Punctuation: "First, close the window; second, turn off the lamp; finally, lock the door."
• Numbers: "The package weighs twelve point six kilograms and arrived on July twenty-first."
• Names/Places: "Gwendolyn photographed the eucalyptus trees outside Ljubljana."
• Technical: "The system runs on three core components that all have to stay in sync."

---

📊 Evaluation & Benchmarks

Rather than using a single unverifiable aggregate score, Inflect v2 reports five distinct metrics to ensure transparency.

Headline Performance: Inflect-Micro-v2

Deep Dive Analysis

1. Human Preference

In a blind community study, Inflect-Micro-v2 achieved a 66.2% preference rate (calculated as 21 wins, 10 losses, and 3 ties, where ties are weighted as half-wins). The study used randomized left/right ordering and hidden system identities.

2. Naturalness vs. Footprint

Using 500 unseen prompts per voice, the UTMOS22 score for Inflect-Micro-v2 was 4.395 (with a 95% bootstrap CI of $4.381–4.408$).

• Note: Competitors like KittenTTS and Piper are represented as two-voice means. Supertonic 3-step is plotted below the main range to avoid scaling issues.

3. Intelligibility (ASR)

The headline Word Error Rate (WER) is the mean of Qwen3-ASR and Nemotron 3.5. Whisper was excluded from the headline because it produced "insertion-heavy hallucinations" on specific Supertonic 8-step clips.

Complete Three-ASR Audit:

---

⚡ Runtime & Deployment

CPU Performance

Tests were conducted on a Hugging Face CPU Upgrade instance:

• Hardware: 8 vCPU, 32 GB RAM
• Config: 4 framework threads, 100 fixed Modern400 prompts.
• Method: The first pass (cache building) was discarded; results are the average of passes two and three.

Comparative Speed Context

The following table provides directional context using a shorter 50-prompt pass.

This is not a perfectly matched speed leaderboard because Inflect used a 100-prompt steady-state run, while others used a 50-prompt pass. Additionally, some competitors utilize optimized ONNX runtimes.