Complete local text-to-waveform speech synthesis under 4M parameters.
Fixed-voice English TTS with deterministic seeds, long-text handling, and CPU or CUDA inference.
A note from Owen
Thanks so much for all the support shown on this project! I built and funded Inflect v2 independently - if this release finds a real audience, I would like to continue the project with a v3, which might include things like more langauges, voices, and stability improvements. If the model is useful to you, leaving a like on Hugging Face genuinely helps more people discover it.
Prepare data, audit train/validation splits, adapt a fixed voice or language, resume training, evaluate checkpoints, and export PyTorch or ONNX packages with the Inflect adaptation toolkit. Adapted quality is experimental and depends on the dataset, frontend, and fluent-speaker evaluation.
Inflect v2 uses one public API across two sizes: Micro prioritizes quality below 10M parameters; Nano prioritizes footprint below 4M.
These are held-out text generations, not reconstructions of training audio. Each transcript is shown exactly as passed to the public frontend.
Test
Exact transcript
Generated audio
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 and places
Gwendolyn photographed the eucalyptus trees outside Ljubljana.
Technical
The system runs on three core components that all have to stay in sync.
Evaluation
No single metric captures TTS quality. Inflect v2 reports human preference, predicted naturalness, multi-ASR intelligibility, complete footprint, and runtime separately rather than compressing them into one unverifiable score.
Community preference ↑
UTMOS22 ↑
Two-ASR semantic WER ↓
Complete FP32 weights ↓
4-thread CPU throughput ↑
63.9%
4.386
4.21%
15.97 MB
10.72× real-time
The headline row always refers to Inflect-Nano-v2. Detailed competitor results and protocol boundaries are kept visible below.
Comparison set. Results include KittenTTS Nano, Piper Low, and Supertonic 3, established compact or local TTS baselines with larger deployable weight footprints than both Inflect releases. Weight sizes are compared at package level, and no single metric is treated as proof of overall superiority.
1. Human blind preference
Community blind listening
Inflect-Nano-v2 recorded a 63.9% preference rate (22 wins · 12 losses · 2 ties) in the final anonymous community study. Systems were hidden, left/right order was randomized, and ties count as half a win. This is descriptive community evidence, not formal MOS.
2. Predicted naturalness versus footprint
Predicted quality versus footprint
The UTMOS22 run used 500 identical unseen prompts per voice. KittenTTS and Piper are equal-weight two-voice means; their observed voice ranges appear as whiskers. Supertonic 3-step is reported below the plotted range rather than flattening every other system.
Inflect-Nano-v2: 4.386 UTMOS22, 95% bootstrap CI 4.372–4.399. UTMOS22 is a learned predictor, not human MOS.
3. Intelligibility on unseen text
Two-ASR semantic WER consensus
The headline score is the equal-weight mean of Qwen3-ASR and Nemotron 3.5 corpus WER for every system. Whisper is excluded consistently from the headline because it produced insertion-heavy hallucinations on a subset of otherwise intelligible Supertonic 8-step clips. It is not deleted: the complete three-ASR evidence remains below.
Open the complete three-ASR audit
Semantic WER across Qwen3-ASR, Nemotron 3.5, and Whisper large-v3
System / voice
Qwen3-ASR ↓
Nemotron 3.5 ↓
Whisper large-v3 ↓
Inflect-Micro-v2
2.52%
5.45%
2.73%
Inflect-Nano-v2
2.79%
5.63%
2.65%
KittenTTS Nano · Bruno
2.15%
3.96%
2.17%
KittenTTS Nano · Hugo
2.39%
3.80%
2.11%
Piper Low · Danny
2.62%
5.60%
2.55%
Piper Low · Ryan
2.81%
5.51%
2.87%
Supertonic 3 · M2 · 3-step
3.03%
6.04%
3.22%
Supertonic 3 · M2 · 8-step
2.05%
3.56%
8.08%
For Inflect-Nano-v2, the individual results are 2.79% Qwen3-ASR, 5.63% Nemotron 3.5, and 2.65% Whisper large-v3. The former three-model mean, 3.69%, is retained only as a descriptive audit value and is not used as the headline score.
Open evaluator robustness and error-category diagnostics
ASR evaluator robustness
Semantic WER by prompt category
These views are diagnostics, not additional leaderboards. They show where the
recognizers disagree and which prompt categories still produce recoverable
transcription errors.
4. CPU runtime
Both Inflect releases synthesize comfortably faster than real time on CPU. The
managed reference run used a Hugging Face CPU Upgrade instance (8 vCPU,
32 GB RAM) with four framework threads, end-to-end text-to-waveform timing,
and 100 fixed Modern400 prompts. Three complete passes were recorded; the first
cache-building pass was excluded and the table pools passes two and three.
Release
Steady-state RTF ↓
Audio / wall time ↑
Inflect-Micro-v2
0.1593
6.28×
Inflect-Nano-v2
0.0933
10.72×
These are package-level results from the public PyTorch runtime, not a claim
that Inflect is the fastest compact TTS system. Hardware, frontend behavior,
framework, compilation, and thread policy all affect small-model measurements.
Open directional compact-system speed context
The same managed CPU and four-thread policy were used for a shorter comparator
pass: the identical 50-prompt prefix, repeated twice. KittenTTS and Piper are
equal-work pooled across their two tested voices.
System
Audio / wall time ↑
Piper Low
31.37×
KittenTTS Nano
13.33×
Inflect-Nano-v2
10.72×
Supertonic 3 · 3-step
10.15×
Inflect-Micro-v2
6.28×
Supertonic 3 · 8-step
4.37×
Because Inflect uses the larger 100-prompt steady-state run while comparator
rows use the shorter 50-prompt confirmation pass, this table is deployment
context rather than a perfectly matched speed leaderboard. Several comparators
also use optimized ONNX runtimes, while the published Inflect benchmark above
uses the canonical PyTorch runtime. The separately released Inflect ONNX path
has not been substituted into those benchmark numbers.
5. Complete weight footprint
Complete deployable model footprint
Voice variants sharing the same weights are merged. Inflect totals include the integrated waveform decoder.
Open the frozen evaluation protocol
Modern400 uses 400 identical unseen English prompts per system: 200 fixed modern/stress prompts plus 200 deterministic FLEURS en_us test prompts.
Exact-text exclusion was checked against 87,362 training transcripts.
All ASR inputs are resampled to 16 kHz and scored with the same disclosed English normalizer.
UTMOS22 uses tarepan/SpeechMOS v1.2.0 on a separate 500-prompt generation set.
Headline intervals use 10,000 bootstrap samples.
The Modern400 corpus SHA-256 is b7504ce2dce44a2da82770a6a5dfd2a034fe17e2113980f8a69663ade417a34c.
Prompts, hypotheses, compressed row-level reports, and summaries ship under evaluation/final/.
Runtime is evaluated separately because framework, thread policy, compilation,
and host load can dominate small-model comparisons.
Choose the right Inflect
Inflect-Nano-v2
Inflect-Micro-v2
Complete parameters
3,966,721
9,356,513
FP32 weights
15.97 MB
37.53 MB
Positioning
Smallest practical footprint
Strongest Inflect v2 quality
24 kHz waveform decoder
Included
Included
Python API and frontend
Same
Same
Inflect-Nano-v2 is the portability-focused member of the family. Both models use the same public API and complete text-to-waveform packaging.
This uses the Hub's version-aware downloader and retrieves the complete
repository. A Git clone also works, but hf download is the recommended path
for ordinary model installation.
Python
python
1from inference import InflectTTS
23tts = InflectTTS(".", device="cpu")4tts.save(5"A small voice can still have something meaningful to say.",6"sample.wav",7 speed=1.0,8 variation=0.667,9 seed=7,10)
The result is a 24 kHz mono float32 waveform. Long input is split at punctuation-aware boundaries, synthesized chunk by chunk, and joined with controlled pauses.
ONNX Runtime
The official verified FP32 export is published separately as
Inflect-Nano-v2-ONNX.
It supports dynamic lengths, CPU/CUDA/DirectML provider selection,
deterministic seeds, and the same long-text wrapper without importing PyTorch:
bash
1git clone https://huggingface.co/owensong/Inflect-Nano-v2-ONNX
2cd Inflect-Nano-v2-ONNX
3python -m pip install -r onnx/requirements.txt
4python onnx/inference_onnx.py \5 --text "The complete model now runs through ONNX Runtime."\6 --output sample-onnx.wav \7 --provider cpu \8 --seed 7
The neural model is split into duration.onnx and decode.onnx; together they
contain the complete learned text-to-waveform path. The English eSpeak-ng
frontend remains CPU-side code. See the
ONNX repository
for graph contracts, provenance, parity measurements, browser deployment, and
re-export instructions.
Release profile
Local runtime
Long-text handling
CPU and CUDA inference through the same Python API and CLI.
Punctuation-aware segmentation with controlled pauses and edge fades.
Repeatable output
Auditable evaluation
Fixed seeds reproduce the same latent sample on the same runtime stack.
Frozen prompts, raw ASR hypotheses, intervals, hashes, and per-system reports are included.
Architecture and parameter budget
Inflect v2 is a parameter-efficient VITS-family end-to-end text-to-waveform generator with an English phoneme frontend, monotonic alignment, stochastic latent synthesis, residual coupling flow, and an integrated alias-reduced neural waveform decoder.
Component
Inflect-Nano-v2
Latent channels
128
Text hidden channels
72
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
Output
24 kHz mono waveform
The release describes the deployable architecture. Private corpus-construction and optimization details are not part of this open-weight package.
Controls, determinism, and long text
Control
Default
Public range
Meaning
speed
1.0
0.5–2.0
Lower is slower; higher is faster.
variation
0.667
0.0–1.0
Lower is steadier; higher samples more latent variation.
seed
0
integer
Repeats the same stochastic sample on the same runtime stack.
Long passages are punctuation-aware chunks, not one unlimited autoregressive pass. Chunk boundaries receive short pauses and edge fades. See docs/API.md for waveform contracts and concurrency notes.
Data, voice, and adaptation status
The release contains one fixed synthetic English voice. The package does not redistribute a real-speaker recording corpus, does not claim the voice as the identity of a real person, and requires no reference audio or external model at inference.
The base release remains inference-first, but a public experimental fixed-voice and language adaptation workflow is now available. A new voice replaces the built-in speaker rather than adding runtime voice cloning. A new language requires owned or licensed speech data, a compatible phoneme frontend, symbol migration, retraining, and fluent-speaker evaluation. Start with the Inflect adaptation toolkit, then review docs/DATA_AND_VOICE.md.
Package map
Path
Purpose
model.pth
Inference-only generator checkpoint
config.json
Architecture and audio configuration; also the Hub download-count query file
inference.py
Public Python API and CLI
inflect_vits_frontend.py
English normalization, phonemization, and punctuation frontend
Separate official FP32 ONNX graphs, torch-free runner, parity report, checksums, and exporter
samples/
Held-out example generations
evaluation/final/
Frozen benchmark prompts, reports, and protocol artifacts
docs/
API, deployment, evaluation, adaptation, and export documentation
release_manifest.json
File sizes and SHA-256 hashes
Limitations
English only, with one fixed male voice. This is 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 pass.
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
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 Apache-2.0 license.
License, integrity, and attribution
Original Inflect code and weights are released under Apache-2.0. Bundled third-party components retain their own notices in THIRD_PARTY_NOTICES.md. release_manifest.json records packaged file sizes and SHA-256 hashes.
Private training scope and contact
Inflect v2 is an open-weight release. Deployable weights, inference code, frontend code, evaluation prompts, and release reports are public. The training corpus-generation pipeline, private filtering infrastructure, and full optimization recipe are not part of the public package.
Owen Song may share additional technical context privately for credible research, collaboration, reproducibility, or deployment inquiries when the request has a clear purpose and does not conflict with licensing or data-provenance constraints.
Discord:b111ue — fastest for informal technical questions
