license: apache-2.0
language:
- en
pipeline_tag: text-to-speech
tags:
- text-to-speech
- speech-synthesis
- local-tts
- cpu
- small-model
- vits
- 24khz
library_name: pytorch
Inflect-Nano-v2
Complete local TTS under 4M parameters.
A complete 24 kHz English text-to-waveform model with its neural waveform decoder included.
Try it in the browser · Inflect-Micro-v2 · Evaluation · Limitations
Inflect-Nano-v2 is the portability-focused member of Inflect v2. The complete deployable system has 3,966,721 parameters; there is no second vocoder download and no hidden large language model at inference time.
Listen First
Every clip below uses held-out text rather than a training-audio reconstruction.
| Prompt type | Audio |
|---|---|
| Conversational | |
| Punctuation | |
| Numbers | |
| Names and places | |
| Technical |
At a Glance
| Inflect-Nano-v2 | |
|---|---|
| Complete deployable parameters | 3,966,721 |
| FP32 model weights | 15.97 MB |
| Sample rate | 24 kHz |
| Voice | 1 fixed English male voice |
| External vocoder | None |
| Release format | PyTorch FP32 |
| Long text | Punctuation-aware chunking |
Inflect v2 is released in two deliberately different sizes. Micro spends its larger budget on waveform capacity and hidden width; Nano prioritizes footprint while preserving the same public runtime and text frontend. Both numbers above count the complete deployable generator, including the waveform decoder.
Human Listening
Inflect-Nano-v2 recorded a 63.9% preference rate (22 wins, 12 losses, 2 ties) in the final anonymous community study. The full study contains 147 pairwise decisions across the published lineup. Ties count as half a win. This is descriptive community evidence, not a formal MOS study; pair coverage and vote counts differ by system.
Evaluation
Quality versus footprint
Intelligibility
Cross-ASR robustness
CPU throughput
Quality and speed
Error analysis
Complete model footprint
| System | Weights | UTMOS22 ↑ | Semantic WER ↓ |
|---|---|---|---|
| Inflect-Micro-v2 | 37.5 MB | 4.395 | 1.93% |
| Inflect-Nano-v2 | 16.0 MB | 4.386 | 2.48% |
| KittenTTS Nano · Bruno | 56.8 MB | 4.217 | 0.97% |
| KittenTTS Nano · Hugo | 56.8 MB | 4.191 | 1.15% |
| Piper · Ryan Low | 63.1 MB | 4.289 | 2.33% |
| Piper · Danny Low | 63.1 MB | 4.195 | 1.81% |
| Supertonic 3 · James · 3-step | 398.1 MB | 2.471 | 692.82% |
| Supertonic 3 · James · 8-step | 398.1 MB | 4.295 | 622.52% |
On the 500-prompt matched set, Inflect-Nano-v2 measured 4.386 UTMOS22 (95% bootstrap CI 4.372–4.399) and 2.48% semantic WER (95% CI 1.95–3.07). UTMOS22 is an automated quality predictor, not human MOS. WER measures transcript recoverability, not naturalness.
Under the prespecified Whisper-large-v3 protocol, Supertonic produced insertion-heavy hypotheses and landed off-scale at 692.82% (3-step) and 622.52% (8-step). The independent wav2vec2 pass scored the same clips at 13.47% and 3.76%, respectively. This large disagreement is reported as ASR-family sensitivity, not collapsed into an aggregate or interpreted as a direct measure of Supertonic intelligibility. Exact hypotheses and per-clip errors ship in the raw reports.
Protocol and raw evidence
- 500 identical unseen English prompts per system; 4,000 generated clips total.
- The set combines a 200-prompt stress corpus with 300 deterministic CMU Arctic phonetic-balance sentences.
- Exact-text exclusion was run against 87,362 training transcripts.
- Headline semantic WER uses Whisper-large-v3 and the disclosed English normalizer; wav2vec2 LV-60K is included as an independent ASR-sensitivity check.
- UTMOS22 uses
tarepan/SpeechMOSv1.2.0. - Headline intervals use 10,000 bootstrap samples.
- Kitten and Piper voices are reported separately; equal-weight family averages are in
evaluation/final/evidence-summary.json. - Supertonic 3-step and 8-step are never pooled.
- Per-clip outputs, category breakdowns, runtime rows, and signal diagnostics are under
evaluation/final/.
Quickstart
pip install -r requirements.txt
python inference.py \
--model-dir . \
--device cpu \
--text "A small voice can still have something meaningful to say." \
--output sample.wav
import sys
from huggingface_hub import snapshot_download
model_dir = snapshot_download("owensong/Inflect-Nano-v2")
sys.path.insert(0, model_dir)
from inference import InflectTTS
tts = InflectTTS(model_dir, device="cpu")
tts.save(
"The complete speech model runs locally.",
"sample.wav",
speed=1.0,
variation=0.667,
seed=7,
)
The packaged runtime accepts speed from 0.5 to 2.0, variation from 0.0 to 1.0, and a deterministic integer seed. Long input is split at punctuation-aware boundaries and concatenated with controlled pauses.
Controls
| Control | Default | Effect |
|---|---|---|
speed |
1.0 |
Speech rate; supported range 0.5–2.0 |
variation |
0.667 |
Stochastic latent variation; lower is steadier, higher is more variable |
seed |
0 |
Reproducible sampling seed; incremented per long-text chunk |
Use a fixed seed when comparing systems. A different seed is a different stochastic sample, not a different checkpoint.
Architecture
Inflect v2 is a parameter-efficient VITS-family, end-to-end text-to-waveform speech generator with a phoneme frontend, monotonic alignment, stochastic latent synthesis, and an integrated neural waveform decoder.
- punctuation-preserving English normalization and stress-marked phonemes
- compact Transformer text encoder and duration model
- latent prior and residual coupling flow
- integrated adversarial waveform decoder at 24 kHz
- inference-only checkpoint with training discriminators and optimizer state removed
| Component | Value |
|---|---|
| Latent channels | 128 |
| Text hidden channels | 72 |
| Text encoder layers / heads | 3 / 2 |
| Feed-forward channels | 384 |
| Flow coupling blocks | 4 |
| Initial decoder channels | 192 |
| Upsample rates | 8, 8, 2, 2 |
| Training segment | 16,384 samples |
The model card describes the deployable architecture. The private corpus-construction and optimization recipe is not part of this open-weight release.
Runtime
Warm end-to-end CPU inference on AMD EPYC 7K62 48-Core Processor, with 192 logical CPUs visible to the isolated process and runtime-default thread settings:
| RTF ↓ | Audio / wall time ↑ | Median latency | p95 latency | Cold load |
|---|---|---|---|---|
| 0.5400 | 1.85× | 2.199 s | 3.850 s | 0.82 s |
Runtime varies with processor, PyTorch build, thread policy, text length, and operating system. Use the raw report rather than extrapolating these measurements to untested devices.
Fine-tuning
The initial release is inference-first. The checkpoint is structurally compatible with voice and language adaptation, but the private training recipe and corpus-generation pipeline are not included. A tested public adaptation toolkit should be treated as a separate release rather than inferred from the inference package. See docs/FINETUNING.md for the supported status and data contract.
Repository Layout
| Path | Purpose |
|---|---|
model.pth |
Inference-only generator checkpoint |
config.json |
Deployable architecture and audio configuration |
inference.py |
Public Python API and CLI |
inflect_vits_frontend.py |
Normalization, phonemization, and punctuation frontend |
runtime/ |
Self-contained model implementation required for inference |
samples/ |
Held-out example generations |
evaluation/final/ |
Final matched benchmark reports and protocol artifacts |
release_manifest.json |
File sizes and SHA-256 integrity hashes |
Limitations
- English only and one fixed male voice. This is not a zero-shot voice-cloning model.
- Unfamiliar phrasing can become flatter or less expressive.
- Numbers, abbreviations, homographs, and uncommon names depend strongly on frontend normalization and context.
- Long passages are synthesized in chunks, so transitions can differ from a single long-form model pass.
- Stochastic variation can change pronunciation and timing; use a fixed seed for reproducibility.
- Learned MOS predictors can reward or penalize spectral traits differently from human listeners.
- Not validated for medical, legal, emergency, or accessibility-critical communication.
Responsible Use
Do not use the included voice to impersonate a real person, deceive listeners, or create fraudulent content. Disclose synthetic speech where the context could otherwise mislead. Users are responsible for applicable laws and the license.
Integrity and License
release_manifest.json records file sizes and SHA-256 hashes. Original Inflect code and weights in this repository are licensed under Apache-2.0; bundled third-party components retain their own notices in THIRD_PARTY_NOTICES.md.
Citation
@software{song2026inflectnanov2,
author = {Owen Song},
title = {Inflect-Nano-v2: Complete Local Text-to-Waveform TTS Under 4M Parameters},
year = {2026},
url = {https://huggingface.co/owensong/Inflect-Nano-v2}
}