Raw capture — raw/inflect-micro-v2.md

url: https://huggingface.co/owensong/Inflect-Micro-v2 title: Inflect-Micro-v2 author: Owen Song date_fetched: 2026-07-29 date_published: 2026

Inflect-Micro-v2 — Complete Model Card

Overview

Inflect-Micro-v2 is a local text-to-speech (TTS) system delivering "complete local text-to-waveform speech synthesis under 10M parameters." It outputs 24 kHz mono audio with a single fixed English male voice, supports deterministic seeds for reproducibility, handles long text via punctuation-aware chunking, and runs on both CPU and CUDA.

Creator: Owen Song (HuggingFace: owensong), an independent developer who built and funded the project without institutional backing.

Architecture

Inflect v2 is a VITS-family end-to-end text-to-waveform generator using:

An English phoneme frontend (eSpeak-ng based)
Monotonic alignment for duration prediction
Stochastic latent synthesis with residual coupling flow (4 flow blocks)
An integrated alias-reduced neural waveform decoder

Component Breakdown (Micro-v2)

Component	Value
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
Training segment	16,384 samples
Output	24 kHz mono

There is also a smaller sibling, Inflect-Nano-v2 (3.97M params, 15.97 MB), which prioritizes minimal footprint over peak quality.

Parameter Count & Footprint

Deployable parameters: 9,356,513
FP32 weight size: 37.53 MB
License: Apache-2.0

Capabilities

Controls

Control	Default	Range	Effect
`speed`	1.0	0.5–2.0	Lower = slower
`variation`	0.667	0.0–1.0	Lower = steadier output
`seed`	0	integer	Reproducible sampling

Long Text Handling

Long passages are split at punctuation boundaries into chunks, synthesized separately, and joined with controlled pauses and edge fades. This is not a single unlimited autoregressive pass.

Output Quality

Demonstrated on held-out text categories including conversational speech, punctuation-heavy text, numbers (spoken-form), names/places, and technical content.

Supported Runtimes

PyTorch (canonical runtime) — CPU and CUDA
ONNX Runtime — published separately as Inflect-Micro-v2-ONNX, supporting CPU/CUDA/DirectML without PyTorch imports, with dynamic lengths and deterministic seeds

Benchmarks

1. Human Blind Preference (Community Study)

Inflect-Micro-v2 achieved a 66.2% preference rate (21 wins, 10 losses, 3 ties) in an anonymous listening study where systems were hidden and left/right order was randomized. Ties counted as half a win.

2. Predicted Naturalness (UTMOS22)

Score: 4.395 (95% bootstrap CI: 4.381–4.408) based on 500 identical unseen prompts. UTMOS22 is a learned predictor, not human MOS.

3. Intelligibility (Multi-ASR Semantic WER)

The headline score is a two-ASR mean (Qwen3-ASR + Nemotron 3.5), excluding Whisper due to insertion-heavy hallucinations on certain competitor clips.

Micro-v2 two-ASR headline: 3.99%

Full three-ASR breakdown:

ASR System	WER
Qwen3-ASR	2.52%
Nemotron 3.5	5.45%
Whisper large-v3	2.73%

Competitor context (two-ASR headline): KittenTTS Nano voices scored ~2.15–2.39% / ~3.80–3.96%; Piper Low voices scored ~2.62–2.81% / ~5.51–5.60%; Supertonic 3 (3-step) scored 3.03% / 6.04%.

4. CPU Runtime (4-thread, 8 vCPU, 32 GB RAM)

Inflect-Micro-v2: 0.1593 RTF → 6.28× real-time
Inflect-Nano-v2: 0.0933 RTF → 10.72× real-time

Speed comparator context (50-prompt pass, not perfectly matched): Piper Low ~31.37×, KittenTTS Nano ~13.33×, Supertonic 3 (3-step) ~10.15×.

5. Complete Weight Footprint

FP32: 37.53 MB for Micro-v2; 15.97 MB for Nano-v2. The integrated waveform decoder is included in both totals.

Comparison Set

The model was benchmarked against KittenTTS Nano, Piper Low, and Supertonic 3 — all compact or local TTS baselines. The developer states that "no single metric is treated as proof of overall superiority."

Training Data

Contains one fixed synthetic English voice — does not redistribute a real-speaker recording corpus
The voice is not claimed as the identity of a real person
The training corpus-generation pipeline and private filtering infrastructure are not part of the public release
"Exact-text exclusion was checked against 87,362 training transcripts"
The release is inference-first, requiring no reference audio or external model at inference

Adaptation (Fine-Tuning)

An experimental fixed-voice and language adaptation toolkit is available at github.com/owenawsong/Inflect/tree/main/finetune. Key constraints:

A new voice replaces the built-in speaker rather than adding runtime voice cloning
A new language requires owned/licensed speech data, a compatible phoneme frontend, symbol migration, retraining, and fluent-speaker evaluation
Adapted quality is described as "experimental" and dataset-dependent

Package Contents

File/Directory	Purpose
`model.pth`	Inference-only generator checkpoint
`config.json`	Architecture/audio config (also Hub download-count query)
`inference.py`	Public Python API and CLI
`inflect_vits_frontend.py`	English normalization, phonemization, punctuation
`runtime/`	Self-contained model implementation
`samples/`	Held-out example generations
`evaluation/final/`	Frozen benchmark prompts, reports, protocol artifacts
`docs/`	API, deployment, evaluation, adaptation, export docs
`release_manifest.json`	File sizes and SHA-256 hashes
`THIRD_PARTY_NOTICES.md`	Third-party component licenses

Limitations (Explicitly Stated)

English only with one fixed male voice — not zero-shot voice cloning
Unfamiliar phrasing can become flatter, less expressive, or less stable
Numbers, abbreviations, homographs, and uncommon names remain frontend- and context-sensitive
Long passages use punctuation-aware chunking; transitions can differ from a native long-form model
Stochastic variation can alter timing and pronunciation (fix the seed for comparisons)
UTMOS22 and ASR scores do not replace controlled human MOS or MUSHRA-style evaluation
Not validated for medical, legal, emergency, or accessibility-critical communication

Responsible Use Guidelines

The developer forbids using the included voice to impersonate a real person, deceive listeners, or create fraudulent content. Synthetic speech disclosure is required where context could mislead. Users bear responsibility for applicable laws and license compliance.

Citation

@software{song2026inflectmicrov2,
  author = {Owen Song},
  title = {Inflect-Micro-v2: Complete Local Text-to-Waveform TTS Under 10M Parameters},
  year = {2026},
  url = {https://huggingface.co/owensong/Inflect-Micro-v2}
}

Community & Contact

Discord: b111ue (fastest for informal questions)
Community server: discord.gg/CVJYedvzvp
Email: owen.aw.song@gmail.com (preferred for professional inquiries)
Downloads (last month): 645
Spaces using this model: 4 (including an official playground and community deployments)
Finetunes: 1 model; Quantizations: 3 models

Intended Use

Complete local (on-device) text-to-waveform speech synthesis — suitable for applications needing a lightweight, self-contained TTS engine with reproducible output and no cloud dependency.