You can deploy this model and start earning money today!

sasha-denisov/
inflect-nano-v2-litert

Name: inflect-nano-v2-litert
Availability: InStock

owensong/Inflect-Nano-v2

finetune

apache-2.0

Views

No views yet

Inflect-Nano-v2 — LiteRT (.tflite) export

A LiteRT export of owensong/Inflect-Nano-v2, a small English text-to-speech model. Upstream ships the canonical PyTorch model and a verified ONNX package, but its docs/EXPORTS.md lists TFLite as "Not released" — this repository fills that gap with a working, target-validated LiteRT conversion for on-device use (Raspberry Pi 5 and similar).

The export mirrors the upstream two-stage decomposition (the same split as the official ONNX duration + decode graphs):

File	Stage	Precision	Size
`inflect_text_encoder.tflite`	text → acoustic distribution	fp32	3.5 MB
`inflect_decoder.tflite`	distribution + noise → waveform	fp32	12.6 MB
`inflect_text_encoder_fp16.tflite`	text → acoustic distribution	fp16	1.8 MB
`inflect_decoder_fp16.tflite`	distribution + noise → waveform	fp16	6.4 MB

Why a separate LiteRT export

Light streaming TTS is often assumed to be non-convertible to LiteRT because of data-dependent output shapes. Exporting the decoder with a fixed chunk size (streaming already operates on fixed-size chunks) sidesteps that: each stage becomes a static graph. Measured on a Raspberry Pi 5: RTF ~0.11, ~225 ms per sentence, bit-exact against the PyTorch reference, streaming within a sentence.

Usage

These are the acoustic graphs only. The non-neural English text frontend (espeak-ng phonemizer, GPL-3.0, run in-process) is not included here — it is not embedded in the graphs, same as the upstream ONNX package. A runnable reference (say.py + frontend) ships with the LiteRT reachy voice-robot sample, which downloads these weights automatically.

python

from ai_edge_litert.interpreter import Interpreter
enc = Interpreter(model_path="inflect_text_encoder.tflite", num_threads=4)
dec = Interpreter(model_path="inflect_decoder.tflite", num_threads=4)
# see say.py in the sample for the full encoder -> duration expansion -> decoder pipeline

Attribution & license

Model, weights, and architecture: © the Inflect-Nano-v2 authors, released under Apache-2.0 at owensong/Inflect-Nano-v2.
This repository redistributes a LiteRT conversion of those weights under the same Apache-2.0 license (see LICENSE). No model weights were retrained or altered beyond format conversion and fixed-shape graph export.
The text frontend (espeak-ng) is GPL-3.0 and is not part of this repository.

sasha-denisov/inflect-nano-v2-litert

Inflect-Nano-v2 — LiteRT (.tflite) export

Why a separate LiteRT export

Usage

Attribution & license

sasha-denisov/
inflect-nano-v2-litert