AXERA-TECH/inflect_nano_v2

Inflect-Nano-v2 for AX650 NPU

High-quality English TTS (VITS architecture) optimized for AX650 NPU3, running entirely on-chip with zero PyTorch dependency at inference time.

• Encoder + Decoder: AX650 NPU (AXMODEL, U16 precision)
• Duration Predictor: ONNX Runtime (onnxruntime)
• Embedding: NumPy lookup
• Real-time Factor: ~0.04x (24× faster than real-time)

🎧 Audio Samples

All samples generated on AX650 NPU, 24kHz.

"Hello and welcome to Inflect Nano V2 text to speech." (3.3s, RTF 0.06x)

"The weather today is sunny with a gentle breeze." (2.4s, RTF 0.07x)

"Once upon a time, there lived a curious robot who loved to sing." (3.7s, RTF 0.05x)

"Neural voice synthesis is the next frontier of artificial intelligence." (3.9s, RTF 0.05x)

"This audio was generated entirely on the AX650 NPU chip." (4.4s, RTF 0.04x)

🚀 Quick Start

bash setup.sh
bash run.sh "Hello world." output.wav

📡 API

# CLI
python -m inflect_tts_sdk --text "Hello" --output out.wav

# OpenAI-compatible Server
python -m inflect_tts_sdk.server --port 8000

📊 Performance

Component	Time	Device
Encoder	2.4 ms	NPU
Decoder	89.6 ms	NPU
DP + Align	14.7 ms	CPU
Total RTF	~0.04x	24× real-time

Measured on AX650 NPU3, 105 tokens → 2.7s audio.

🧩 Model

Component	Size	Input	Output
Encoder	1.0 MB	x_emb [1,72,200]	m_p, logs_p, x, x_mask
Decoder	5.6 MB	z_p [1,128,500]	waveform [1,1,128000]

• Quantization: Encoder Conv U16 / MatMul INT8, Decoder Conv+ConvTranspose U16
• Parameters: 4.0M
• Sample rate: 24 kHz
• Source: owensong/Inflect-Nano-v2
• Compiled: Pulsar2 6.0-lite, AX650 NPU3

📁 File Structure

• models/ — AXMODEL, dp.onnx, emb_weight.npy
• python/inflect_tts_sdk/ — Python SDK (CLI + OpenAI Server)
• cpp/inflect_tts_sdk/ — C++ SDK (CLI + Server)
• samples/ — audio samples