Skip to content
Audio

M5-TTS.AXERA

accelerated text-to-speech and voice cloning on AX650 NPU

2 min read

The model is an optimized version of the compact M5-TTS-50M-VoiceClone model (50M parameters), adapted to run on hardware accelerators and boards featuring the AX650 NPU. The system handles zero-shot (without prior fine-tuning) text-to-speech synthesis and voice cloning across six languages (including Chinese and English) with streaming support. The output is a high-quality audio file or stream with minimal latency. The development targets embedded systems and edge AI, where high operating speeds on specialized hardware are required without the need for powerful GPUs.

Key features

  • Hardware acceleration: Full execution on AX650 NPU chips (6 axmodel files, Pulsar2 8.0 quantization with U16/S8 and SmoothQuant).
  • Voice cloning: Support for zero-shot speech generation from a short reference audio recording (up to 4.8 seconds).
  • Multilingual support: Support for Chinese, English, German, French, Japanese, and Korean.
  • Low latency: Streaming mode delivers the first audio chunk (TTFA) starting from 0.37 seconds in the C++ implementation.

What you can do

  • Run on an AX650 board: Deploy a local voice assistant or auto-responder system on embedded hardware.
  • Clone voices: Synthesize phrases in different languages using a short reference voice sample and precise transcript text.
  • Use streaming mode: Integrate real-time speech generation with a minimal RTF (0.48–0.58).
  • Build a C++ or Python app: Use ready-to-run scripts and SDKs to embed the model into your own projects.

Example

import sys; sys.path.insert(0, "python")
from m5_ax import M5AX, write_wav

tts = M5AX("models", "models/assets")
wav = tts.synthesize("你好,欢迎使用我们的语音合成模型。",
                     "samples/aslp_default_voice.wav",
                     "猪携带该疾病,然后通过蚊子传染给人类。")
write_wav("out.wav", wav, tts.sr)
print(tts.last_timing)   # ttft / ttfa / rtf
Read moreDetails
Supported platforms: AX650N (tested), other AX650 boards.
Performance: RTF 0.48–0.58 in C++ (for long phrases), offline and streaming modes.
Hardware requirements: AX650 board with AXEngine 2.12 SDK and Pulsar2 8.0.
License: Apache-2.0.
Discuss Telegram 0

Will it run?

Working out the memory needs

The bot is still collecting this model's weight sizes; the estimate will appear here in a few minutes. Meanwhile you can rent a GPU with headroom.

Massed ComputeOpen Massed Compute

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # HY-2012/M5-TTS.AXERA
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download HY-2012/M5-TTS.AXERA --local-dir ~/models/M5-TTS.AXERA
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to M5-TTS.AXERA
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of M5-TTS.AXERA (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

HY-2012/M5-TTS.AXERA
Author
HY-2012
Category
Audio
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments