---
language:
  - en
pipeline_tag: text-to-speech
license: other
tags:
  - text-to-speech
  - tts
  - speech
  - local
  - cpu
  - onnx
  - small-model
  - experimental
---

# Inflect-Micro-v2

> **Private release preview.** The selected male checkpoint is frozen for final
> packaging. Female weights, inference files, samples, and release evaluations
> will be added before the repository is made public.

**A complete 24 kHz English text-to-speech model in 9.36 million inference
parameters. The waveform decoder is included.**

Inflect-Micro-v2 is the quality-focused member of the Inflect v2 family. It is
being built for local assistants, games, embedded applications, and research on
high-quality TTS under a strict 10M-parameter limit.

[Inflect-Nano-v2](https://huggingface.co/owensong/Inflect-Nano-v2) -
[Inflect-Nano-v1](https://huggingface.co/owensong/Inflect-Nano-v1)

## Current Release Candidate

| Property | Male candidate | Female candidate |
| --- | ---: | ---: |
| Selected checkpoint | 180K polish | Pending |
| Total deployable parameters | **9,356,513** | Target: same architecture |
| Canonical FP32 size | **37.47 MB** | Pending |
| FP16 size | **18.75 MB** | Pending validation |
| Sample rate | **24 kHz** | **24 kHz planned** |
| Input | English text | English text |
| Output | Waveform | Waveform |
| External vocoder required | No | No |

The parameter count describes the inference-only text-to-waveform model. It does
not include optimizer state, discriminators, or training-only posterior paths,
and it does not hide a separate vocoder.

## Architecture

Inflect-Micro-v2 uses a compact VITS-derived, single-stage text-to-waveform
backbone configured specifically for the sub-10M regime:

- a normalized English phoneme frontend with stress and preserved punctuation
- a deterministic duration predictor for compact, stable deployment
- monotonic alignment between phonemes and acoustic frames
- a residual-coupling latent flow and posterior encoder during training
- an integrated HiFi-GAN-style waveform decoder; no external vocoder is needed
- multi-resolution STFT supervision during waveform training

The selected checkpoint does not use the separate experimental HierJoint or
HierNAR prosody-planner branches evaluated during development. Those models are
not part of this release candidate.

### Core Configuration

| Setting | Value |
| --- | ---: |
| Latent channels | 192 |
| Text hidden channels | 96 |
| Feed-forward channels | 768 |
| Text encoder layers | 3 |
| Attention heads | 2 |
| Posterior encoder layers | 3 |
| Initial waveform-decoder channels | 320 |
| Segment size | 16,384 samples |
| Upsampling | 8 x 8 x 2 x 2 |

## Evaluation Preview

These measurements are **provisional development results**, not final release
claims. The full matrix will be rerun from the exact downloadable male and
female artifacts with published prompts, hashes, generation settings, raw
outputs, and confidence intervals.

| Metric | Current male 180K | Protocol |
| --- | ---: | --- |
| UTMOS22 | **4.370** | 200 matched prompts; automated MOS predictor |
| Semantic WER | **3.55%** | Whisper-large-v3; 48 difficult OOD prompts |
| FP32 artifact size | **37.47 MB** | Inference-only Safetensors |

The current semantic-WER confidence interval overlaps the compact Piper and
Kitten baselines. The supported interpretation is **competitive intelligibility
on this prompt set**, not a definitive win.

An earlier public listening study collected 196 pairwise decisions, but it did
not record the precise Micro checkpoint in its metadata. It will not be used as
evidence for this 180K release candidate. A new exact-artifact study is planned.

## Planned Release Files

```text
weights/
  male/model.fp32.safetensors
  male/model.fp16.safetensors
  female/model.fp32.safetensors
  female/model.fp16.safetensors
onnx/
  male/model.onnx
  female/model.onnx
examples/
evaluation/
voices.json
inference.py
app.py
```

FP32 Safetensors will be the canonical quality artifact. FP16 and ONNX will only
be released after waveform-level parity and listening checks. INT8/GGUF formats
are not planned for launch because they require their own audio-quality gate.

## Intended Use

- local and offline speech generation
- assistants, games, prototypes, and efficient deployment research
- applications where model size and CPU execution matter

## Known Limitations

- English only
- fixed voices; this is not zero-shot voice cloning
- unfamiliar phrasing can become flatter or more artifact-prone
- punctuation and prosody are learned rather than guaranteed
- numbers, abbreviations, and homographs depend on the text frontend
- long passages should be sentence-chunked by the runtime
- synthetic teacher speech can transfer teacher-specific biases
- not validated for critical medical, legal, emergency, or accessibility use

## Training Data and License

The model was trained primarily on synthetic English speech generated by a
larger teacher TTS system. The final release will disclose corpus size, teacher
configuration, filtering, train/test separation, and reference-voice
provenance.

Repository code is Apache 2.0 unless otherwise noted. The final model-weight and
synthetic-data terms are still under review, so this preview uses
`license: other`. Upstream code, teacher-model, dataset, and voice terms are not
automatically relicensed by this repository.

## Responsible Use

Do not use the included voices to impersonate a real person, deceive listeners,
or create fraudulent content. Disclose synthetic speech where context could
otherwise mislead.

## Release Gate

Before this repository becomes public:

- [ ] train and select the female checkpoint
- [ ] upload inference-only FP32 artifacts and SHA-256 hashes
- [ ] validate FP16 and ONNX parity
- [ ] run exact-artifact WER, UTMOS22, runtime, and failure audits
- [ ] complete an exact-checkpoint blind listening study
- [ ] publish representative, average, OOD, long, and failure samples
- [ ] verify installation and inference from a clean machine
- [ ] finalize model and dataset licensing