Downloads · 30 days
35
17% of all-time downloads
Arm/qwen3-tts-0-6b-int8-onnx-graviton-g4
qwen3-tts-0-6b-int8-onnx-graviton-g4 is a text-to-speech model from Arm. Use it when you need text read aloud. It is set up for onnxruntime. The card lists the license as apache-2.0.
Qwen3-TTS-0.6B zero-shot voice-cloning text-to-speech optimized as a mixed-precision INT8 ONNX cascade for Arm-based Cloud CPU systems.
Downloads · 30 days
35
17% of all-time downloads
All-time downloads
206
Public
Repo size
3.8 GB
Likes
0
Public
Click a slice to open those files.
.onnx2.5 GB · 67%
From the Hugging Face model README
Qwen3-TTS-0.6B zero-shot voice-cloning text-to-speech optimized as a mixed-precision INT8 ONNX cascade for Arm-based Cloud CPU systems.
This repository contains an Arm-optimized version of Qwen/Qwen3-TTS-12Hz-0.6B-Base for zero-shot voice-cloning text-to-speech. The model is provided as an ONNX cascade of 11 units executed by ONNX Runtime, targeting Cloud CPU systems. Given input text and a short reference clip, it synthesizes the text in the reference speaker's voice at 24 kHz.
Each cascade unit carries the precision it tolerates best: FP16 for the talker backbone and decode, FP32 for the speaker encoder and sub-talker prefill, and INT8 for sub-talker decode, sub-talker heads, and the codec decoder. Quantizing the sub-talker decode loop — the autoregressive step that dominates wall time — is the decisive change; the talker itself babbles under INT8 and is kept at FP16.
This version is intended to demonstrate efficient inference on Arm-based platforms while preserving the original model's intended behavior. Arm has evaluated this model on the English subset of the TTS Multilingual test set and measured performance on a representative evaluation target.
Key results
| Area | Result |
|---|---|
| Model format | ONNX cascade (11 units, mixed precision) |
| Target device class | Cloud CPU |
| Reference device | AWS Graviton4 r8g.xlarge (Neoverse-V2, 4 vCPU, Linux Ubuntu 24.04) |
| Primary performance result | 27.67 s p50 end-to-end synthesis latency, 1.79x faster than the FP32 ONNX baseline |
| Accuracy result | WER 1.5% (Whisper-small), Speaker SIM 0.9400 (WavLM) |
| Size / memory result | 2413.20 MB cascade size (naive sum), 1.95x smaller than FP32 |
| Field | Value |
|---|---|
| Original model | Qwen3-TTS-12Hz-0.6B-Base |
| Original source | Hugging Face |
| Original developer | Alibaba Cloud (Qwen team) |
| Original model card | Qwen/Qwen3-TTS-12Hz-0.6B-Base |
| Original license | Apache-2.0 |
This is not a single-file model. Every ONNX unit below, plus text_path_weights.npz and text_path_config.json, must be present for the cascade to run.
| File | Description |
|---|---|
talker_backbone_fp16.onnx | Talker backbone (prefill), FP16 |
talker_decode_fp16.onnx | Talker autoregressive decode step, FP16 |
sub_talker_backbone_fp32.onnx | Sub-talker backbone (prefill), FP32 |
sub_talker_decode_int8.onnx | Sub-talker autoregressive decode step, INT8 |
sub_talker_heads_int8.onnx | Sub-talker residual-codebook heads, INT8 |
codec_decoder_int8.onnx | Codec tokens to 24 kHz waveform, INT8 |
speaker_encoder_fp32.onnx | Reference-speaker embedding, FP32 |
talker_embed_fp32.onnx | Talker embedding table, FP32 |
subtalker_embed_fp32.onnx | Sub-talker embedding table, FP32 |
codec_head_fp32.onnx | Codec head, FP32 |
text_path_weights.npz | Text-path weights evaluated in NumPy |
text_path_config.json | Text-path and cascade configuration |
onnx_cascade_driver.py | Cascade driver that wires the units together |
example.py | Minimal inference example |
pyproject.toml | Pinned runtime dependencies for example.py, resolved with uv |
uv.lock | Locked dependency resolution for pyproject.toml |
.python-version | Python release the environment was locked and validated on |
config.yaml | Model I/O contract used by the example |
metadata.yaml | Model metadata used by Model Garden |
sample_input.wav | Example reference clip whose speaker is cloned |
sample_output.wav | Example output produced by the optimized model |
prediction.json | Example run summary produced by the optimized model |
benchmarks/ | FP32 baseline and Arm-optimized benchmark records |
Performance was measured on the reference configuration below. Results make the optimization reproducible but do not guarantee identical performance on every Arm-based system.
Reference configuration
| Field | Value |
|---|---|
| Device / platform | AWS Graviton4 r8g.xlarge |
| CPU / accelerator | Neoverse-V2 (4 vCPU), CPU |
| OS | Linux — Ubuntu 24.04 |
| Runtime | ONNX Runtime 1.26.0 |
| Backend / delegate | CPUExecutionProvider (MLAS NEON, KleidiAI), 4 intra-op threads, graph optimization level all |
| Batch size | 1 |
| Precision | Mixed — FP16 talker, FP32 speaker encoder and sub-talker prefill, INT8 sub-talker decode, sub-talker heads, and codec decoder |
| Runs | 0 warmup + 99 measured (one greedy synthesis per evaluation sample) |
Performance results
| Metric | Original / baseline | Arm-optimized | Improvement |
|---|---|---|---|
| Mean latency | 47.09 s | 27.25 s | 1.73x |
| p50 latency | 49.55 s | 27.67 s | 1.79x |
| p90 latency | 59.17 s | 34.35 s | 1.72x |
| Per-step decode latency (length-normalized) | 451.7 ms | 261.7 ms | 1.73x |
| Model size (naive sum) | 4709.28 MB | 2413.20 MB | 1.95x smaller |
| Peak memory | 16184.9 MB | 15741.6 MB | 1.03x less |
Latency is end-to-end voice-clone synthesis for one utterance across the full cascade, and scales with text length; the per-step decode row is the length-normalized autoregressive step. The baseline column is the FP32 ONNX cascade run through the same ONNX-only driver, not the original PyTorch implementation.
Accuracy was evaluated using the same text, reference clips, and evaluation protocol for both the baseline and the optimized cascade.
Evaluation setup
| Field | Value |
|---|---|
| Dataset | TTS Multilingual test set (English subset) |
| Split | validation |
| Number of samples | 99 |
| Metric(s) | Word Error Rate (openai/whisper-small transcription), Speaker SIM (microsoft/wavlm-base-plus-sv) |
| Evaluation runtime | ONNX Runtime |
Accuracy results
| Metric | Original / baseline | Arm-optimized | Change |
|---|---|---|---|
| WER (Whisper-small) | 1.2% | 1.5% | +0.3 pp |
| Speaker SIM (WavLM) | 0.9419 | 0.9400 | -0.0019 |
No sample degenerated into runaway generation across the 99 evaluated utterances. Accuracy was measured using the evaluation setup described above. Users should re-evaluate the model on their own data before production use.
Arm optimized this model for efficient inference on Arm-based platforms using a hardware-aware conversion and validation flow.
For this release, Arm used:
| Optimization area | Applied? | Notes |
|---|---|---|
| Model conversion | Yes | Exported to an 11-unit ONNX cascade driven by onnx_cascade_driver.py, with the text path evaluated in NumPy |
| Quantization | Yes | INT8 post-training quantization, symmetric per-channel weights — dynamic for sub-talker decode and heads, statically calibrated (16 samples) for the codec decoder; FP16 talker, FP32 speaker encoder and sub-talker prefill |
| Runtime/backend selection | Yes | ONNX Runtime CPUExecutionProvider with MLAS NEON and KleidiAI kernels |
| Graph/runtime compatibility updates | Yes | Performed as part of the ONNX export and per-unit precision cast |
| Accuracy validation | Yes | Compared against the original model or published baseline |
| Performance validation | Yes | Measured on the reference Arm platform |
The goal of this process is to improve deployment characteristics such as latency, memory use, model size, and runtime compatibility while preserving the model's intended behavior. Detailed conversion scripts, calibration configuration, or backend-specific implementation details may be provided separately where appropriate.
Dependencies are declared in pyproject.toml, which ships with this repository. Resolve and install them into a local virtual environment with uv:
uv python install
uv sync --frozen
No distribution packages are required: the locked manylinux wheels provide the native libraries the example needs, including libsndfile for soundfile.
Check the declared runtime capability before inference:
uv run python -c 'import onnxruntime as ort; assert "CPUExecutionProvider" in ort.get_available_providers()'
uv run example.py
The example synthesizes one utterance and prints a JSON summary. It writes no files by default; pass an explicit output path to save the generated waveform:
uv run example.py --output /tmp/qwen3-tts.wav
The committed result files are prediction.json and sample_output.wav. An ordinary example run does not read, compare, or overwrite them.
| Property | Value |
|---|---|
| Input shape | Text tokenized to input_ids, int64 [1, T]; reference waveform, float32 mono [N] |
| Input type | UTF-8 text and a 24 kHz mono float32 waveform |
| Input range | Reference waveform in [-1.0, 1.0] |
| Preprocessing | Wrap the text in the Qwen3-TTS assistant prompt and tokenize it to input_ids; resample the reference clip to 24 kHz mono with no further normalization |
| Property | Value |
|---|---|
| Output shape | Mono waveform [M], where M scales with the number of autoregressive decode steps |
| Output type | float32 in [-1.0, 1.0] at 24 kHz |
| Postprocessing | Autoregressive talker decode of the codec token stream to EOS, sub-talker residual-codebook decode over 16 code groups, then codec decoding to a 24 kHz waveform |
This model is intended for developers evaluating zero-shot voice-cloning text-to-speech workloads on Arm-based platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.
example.py imports the committed onnx_cascade_driver.py from this directory to wire the units together.Original Model: Qwen3-TTS-12Hz-0.6B-Base by Alibaba Cloud (Qwen team) - Repository
Optimization/conversion: Arm-Optimized version for execution on Arm-based platforms.
Converted/optimized by: Arm
License: The Original Model and the Optimized Model are subject to Apache-2.0.
This repository contains a converted or optimized version of the Original Model (the “Optimized Model”). The Original Model has been converted or optimized as described above for execution on Arm-based platforms.
No retraining or fine-tuning of the Original Model was performed as part of the conversion or optimization. The conversion or optimization was not intended to change the Original Model’s behavior or intended use.
For information about the Original Model, including its development, training data, intended uses, limitations and other relevant information, please refer to the Original Model repository. Information in that repository was provided by the original developer or other third parties and, unless expressly stated otherwise, has not been independently verified by Arm.
Use of the Original Model and the Optimized Model is subject to the applicable licenses, usage restrictions and other terms identified above and in the relevant repositories. Publication of the Optimized Model does not grant any rights beyond those provided under the applicable license terms.
You are responsible for reviewing those terms and ensuring that your use of the Original Model and the Optimized Model is permitted.
The Optimized Model is provided as a reference implementation to demonstrate and evaluate execution and performance on Arm-based systems. It is not a production-ready or supported solution.
Arm’s publication of the Optimized Model does not constitute an endorsement or certification of the Original Model or a representation that the Optimized Model is suitable for production use or any particular purpose.
To the fullest extent permitted by applicable law (i) the Optimized Model is provided “as is.” Arm makes no representations or warranties that the Original Model, the Optimized Model or their outputs are accurate, safe, secure, non-infringing, legally compliant, suitable for production use or fit for any particular purpose; and (ii) Arm will not be liable for any loss or damage arising from or in connection with the Optimized Model, its use or its outputs.
You are responsible for independently evaluating the Optimized Model, its outputs and its suitability for your intended use, including compliance with applicable legal, regulatory, safety and security requirements.
Arm does not commit to provide ongoing support, maintenance or updates for the Optimized Model. Any use of or reliance on the Optimized Model or its outputs is at your own risk.