Inflect-Nano is an ultra-small local English TTS project. Inflect-Nano-v1 is released on Hugging Face as a complete 4.63M-parameter text-to-waveform stack; Inflect-Nano-v2 is active research toward stronger 4M and 10M variants.

v1 Release · Quickstart · Architecture · v2 Research · Evaluation · Publishing

---

Inflect-Nano-v1

Inflect-Nano-v1 is a tiny English text-to-speech model with 4.632M total inference parameters, including its vocoder.

Released model:

• owensong/Inflect-Nano-v1 on Hugging Face

The goal is not to beat large TTS systems. The goal is to prove how far a complete local text-to-waveform stack can be pushed at an extremely small size.

Highlights

• 4.63M total inference parameters
• Includes the vocoder
• 24 kHz audio
• Single English male voice
• Local PyTorch inference
• Built for tiny-model experiments, local assistants, embedded demos, and efficient inference research

Model Size

Part	Parameters
Acoustic model	3.465M
Vocoder generator	1.167M
Total inference stack	4.632M

Released model files:

weights/inflect_nano_v1_acoustic.pt
weights/inflect_nano_v1_vocoder.pt

Quickstart

Use the Hugging Face release for the runnable v1 model:

git clone https://huggingface.co/owensong/Inflect-Nano-v1
cd Inflect-Nano-v1
pip install -r requirements.txt
python inference.py --text "Wait, are you actually being for real now?" --out sample.wav

CPU:

python inference.py --device cpu --text "Please say neighborhood clearly." --out sample_cpu.wav

Simple controls:

python inference.py \
  --text "The appointment moved to 1:25." \
  --length-scale 1.03 \
  --pitch-scale 1.00 \
  --energy-scale 1.00 \
  --out sample_controlled.wav

Local Gradio demo:

python app.py

Architecture

Inflect-Nano-v1 uses a compact non-autoregressive acoustic model plus a small waveform generator.

flowchart LR
  A["Text"] --> B["English text frontend"]
  B --> C["Compact FastSpeech-style acoustic model"]
  C --> D["80-bin mel spectrogram"]
  D --> E["Small Snake HiFi-GAN-style vocoder"]
  E --> F["24 kHz waveform"]

Loading

Main v1 settings:

Setting	Value
Sample rate	24 kHz
Mel bins	80
Acoustic hidden size	168
Encoder layers	5
Decoder layers	6
Vocoder upsample rates	8, 8, 2, 2

The acoustic model predicts duration, pitch, energy, and brightness before decoding mel frames. The vocoder is a small Snake-activation HiFi-GAN-style generator trained for 24 kHz reconstruction.

Good For

• Tiny local TTS experiments
• Offline assistant prototypes
• Efficient inference research
• Embedded speech demos
• Browser/WASM-style exploration
• Baseline comparisons for sub-5M TTS work

Not Good For

• Production narration
• Accessibility-critical output
• Voice cloning
• Multilingual speech
• High-fidelity audiobook generation
• Matching large modern TTS systems

Limitations

Inflect-Nano-v1 is a very small experimental model. It can sound robotic, buzzy, or unstable, especially on difficult unseen text. Long prompts and unusual phrasing are less reliable, and the vocoder is a clear quality bottleneck.

Use v1 as a tiny-model research/demo release, not as a production TTS engine.

Inflect-Nano-v2

Inflect-Nano-v2 is the active research track. It is not released yet.

Current direction:

• two planned sizes: approximately 4M and 10M total inference parameters
• single-voice English first, with cleaner finetuning paths later
• teacher-distilled data from larger TTS systems
• stronger acoustic model experiments around prior-plus-residual CFM
• vocoder bakeoffs around compact HiFi-GAN, iSTFTNet, and source-filter variants
• objective checks plus listening tests before any release claim

The v2 rule is strict: do not trust training loss by itself. A candidate has to survive reconstruction tests, unseen-text listening, objective diagnostics, and a clear comparison against v1.

Evaluation

Inflect is evaluated with both human listening and objective checks.

flowchart TD
  A["Generate fixed prompt set"] --> B["Blind A/B listening"]
  A --> C["Objective audio checks"]
  B --> D["Accept, reject, or retry experiment"]
  C --> D
  D --> E["Document result"]

Loading

Tracked checks include:

• pronunciation and text coverage
• speaker consistency for single-voice releases
• pacing and duration stability
• skipped or invented words
• leading clicks, long silences, and internal dropouts
• loudness jumps
• real-time factor
• vocoder artifacts

See docs/EVALUATION.md.

Repository Map

Path	Purpose
inflect/	Inflect-native modules and extension experiments.
scripts/	Dataset generation, training, rendering, evaluation, and release tooling.
inflect_asr/	Side project for small ASR and teacher-label pipelines.
voice-encoder/	Voice conditioning and paralinguistic research.
docs/	Architecture, roadmap, evaluation, media kit, and release notes.
examples/	Lightweight public examples and sample assets.
assets/	README and media-kit visuals.

Local generated outputs, checkpoints, reference voices, virtual environments, and third-party checkouts are intentionally excluded from GitHub.

Publishing

GitHub is the source, docs, and experiment-planning home. Hugging Face is the release home for runnable model weights and model cards.

Current public release:

• owensong/Inflect-Nano-v1

Future v2 releases should not replace v1 until they have better listening results, cleaner diagnostics, and a reproducible release package.

License

Repository code and documentation are licensed under Apache 2.0 unless otherwise noted.

Generated datasets, reference voices, model checkpoints, and third-party components may have separate terms. See LICENSE, PUBLISHING.md, and SECURITY.md.