miotts-0.6b-edgellm-aarch64

smajji/miotts-0.6b-edgellm-aarch64

miotts-0.6b-edgellm-aarch64 is a text-to-speech model from smajji. Use it when you need text read aloud. It is set up for tensorrt-edgellm. The card lists the license as apache-2.0.

TensorRT-Edge-LLM engine + codec assets for running MioTTS-0.6B (a Qwen3-0.6B-based TTS LLM) on an aarch64 edge device (Jetson Orin) with TRT 10.13 + CUDA 12.1.

Downloads · 30 days

Access

Public

Updated Sep 6, 2026

Repo size

1.1 GB

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.engine573 MB · 53%

At a glance

Task: Text-to-Speech
Library: tensorrt-edgellm
License: apache-2.0
Access: Public
Created: Sep 6, 2026
Updated: Sep 6, 2026
SHA: 7fda3e4b

Base models

Aratako/MioTTS-0.6B

Model card

From the Hugging Face model README

MioTTS-0.6B (TensorRT-Edge-LLM) — aarch64 engine bundle

TensorRT-Edge-LLM engine + codec assets for running MioTTS-0.6B (a Qwen3-0.6B-based TTS LLM) on an aarch64 edge device (Jetson Orin) with TRT 10.13 + CUDA 12.1.

Pipeline: text → MioTTS talker (TRT, INT4-AWQ + FP8 embedding) → <|s_N|> speech tokens → MioCodec decoder (ONNX Runtime) → iSTFT → 24kHz wav.

Files

Path	Contents
`engines/talker/*`	INT4-AWQ + FP8-embedding talker engine dir (`llm.engine`, `embedding.safetensors`, `config.json`, tokenizer, `processed_chat_template.json`)
`codec/miocodec_decoder.onnx`	MioCodec wave-decoder graph (content tokens + global embed → magnitude/phase); iSTFT done in the runtime
`codec/config.yaml`	MioCodec-25Hz-24kHz config
`presets/*.bin`	128×f32 global-embedding voice presets (en/jp female/male)
`samples/*.wav`	24kHz sample outputs

Runtime notes

Talker model: model_type: qwen3, 28 layers, hidden 1024, vocab 164480 (speech tokens 151669..164468, index = id − 151669).
Talker engine built with llm_build --maxInputLen 1024 --maxKVCacheCapacity 1024.
The decoder ONNX (content_token_indices int64 [B,T], global_embedding f32 [B,128] → magnitude/phase [B,961,S], S = 2·T) runs in ONNX Runtime; use ≥1.20 (the on-board 1.16 lacks Trilu). iSTFT (n_fft 1920, hop 480) is done on device (cuFFT / host).
Codec config: sample_rate 24000, use_wave_decoder: true, wave_upsample_factor: 2, wave_decoder_dim: 512, istft_padding: same.

See tools/miotts/README.md in the source tree for the full port runbook, and run_mio.sh for a server/CLI launcher.

Task: Text-to-Speech
Library: tensorrt-edgellm
License: apache-2.0
Created: Sep 6, 2026
Updated: Sep 6, 2026