Arm/qwen3-tts-0-6b-int8-litert-vivo-x300

Qwen3-TTS-12Hz-0.6B-Base optimized for Arm-based mobile CPUs with SME2

Qwen3-TTS-12Hz-0.6B-Base, a zero-shot voice-cloning text-to-speech model, converted to a LiteRT .tflite cascade for Arm-based mobile CPUs with SME2.

Summary

This repository contains an Arm-optimized version of Qwen/Qwen3-TTS-12Hz-0.6B-Base for autoregressive text-to-speech with voice cloning. The model is provided as a LiteRT .tflite cascade running on the litert CPU runtime, targeting Mobile CPU systems.

The model does not deploy as a single file. It is a 10-unit cascade — talker backbone/decode/embed, sub-talker backbone/decode/heads/embed, codec head, codec decoder, and speaker encoder — driven step by step at runtime. Given input text and a short reference clip, it synthesizes the text in the reference speaker's voice at 24000 Hz.

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 tts_multilingual (English subset) and measured performance on a representative evaluation target.

Key results

Area	Result
Model format	LiteRT .tflite
Target device class	Mobile CPU
Reference device	vivo X300 (8-core Arm CPU (1x C1-Ultra + 3x C1-Premium + 4x C1-Pro), android 16)
Primary performance result	287030 ms p50 end-to-end synthesis latency, 1.60x faster than the FP32 cascade; 0.393 decode steps per second
Accuracy result	Normalised WER 1.11 (openai/whisper-small), speaker similarity 0.9207 (microsoft/wavlm-base-plus-sv)
Size / memory result	5722.9 MB total cascade, 1.63x smaller than the FP32 cascade

Original model

Field	Value
Original model	Qwen/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

Model files

File	Description
example.py	Minimal voice-clone inference example
config.yaml	Cascade manifest: per-unit precision, dispatch rule, and I/O contract
metadata.yaml	Model metadata used by the example or Model Garden
pyproject.toml	Pinned runtime dependencies and deployment metadata
uv.lock	Locked dependency versions, sources, and hashes for uv sync --frozen
benchmarks/	FP32 baseline and Arm-optimized benchmark records

Performance

Performance was measured on the reference configuration below. Results are intended to make the optimization reproducible but do not guarantee identical performance on every Arm-based system.

Reference configuration

Field	Value
Device / platform	vivo X300
CPU / accelerator	8-core Arm CPU (1x C1-Ultra + 3x C1-Premium + 4x C1-Pro), 4.21 GHz peak, 16 GB system memory, CPU execution
OS	android 16
Runtime	litert — LiteRT 0.9.0 (on-device C++ runtime)
Backend / delegate	XNNPACK, KleidiAI
Batch size	1
Precision	mixed (int8 codec_decoder + sub_talker_decode/heads, fp32 elsewhere), PTQ-weight-only, 8-bit weights, symmetric, per-channel
Runs	0 warmup + 17 measured (paired FP32 baseline / optimized)

Performance results

Metric	Original / baseline	Arm-optimized	Improvement
p50 end-to-end latency	460630 ms	287030 ms	1.60x faster
p90 end-to-end latency	590950 ms	345530 ms	1.71x faster
p99 end-to-end latency	731300 ms	402270 ms	1.82x faster
Mean per-step decode latency	4443.5 ms	2542.9 ms	1.75x faster
p50 per-step decode latency	4221.1 ms	2464.7 ms	1.71x faster
Decode steps per second	0.225	0.393	1.75x
Real-time factor (RTFx)	0.0212	0.0451	2.13x
Model size	9334.8 MB	5722.9 MB	1.63x smaller
Peak memory	10926.9 MB	10812.6 MB	1.01x less
Average memory	4800.0 MB	8325.3 MB	0.58x — higher, see Additional notes

Accuracy

Accuracy was evaluated using the same preprocessing, input resolution, and evaluation protocol described below. Where possible, the optimized model is compared against the original model under the same evaluation conditions.

Evaluation setup

Field	Value
Dataset	tts_multilingual
Split	English subset
Number of samples	17 paired on-device runs (FP32 baseline / Arm-optimized)
Metric(s)	Normalised WER (transcribed with openai/whisper-small), speaker similarity (microsoft/wavlm-base-plus-sv)
Evaluation runtime	litert

Accuracy results

Metric	Original / baseline	Arm-optimized	Change
Normalised WER	0.92	1.11	+0.19
Speaker similarity	0.9218	0.9207	-0.0011

Accuracy was measured using the evaluation setup described above. Users should re-evaluate the model on their own data before production use.

Arm optimization approach

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	Each cascade unit traced and converted to LiteRT .tflite
Quantization	Yes	Post-training weight-only INT8: 8-bit weights, 32-bit activations, symmetric, per-channel, no activation calibration. Applied to codec_decoder, sub_talker_decode (dynamic) and sub_talker_heads; all other units left at FP32
Runtime/backend selection	Yes	LiteRT CPU runtime with XNNPACK and KleidiAI
Graph/runtime compatibility updates	Yes	Power-to-multiply operator rewrite for TFLite op coverage, a decode-step cache-position fix for the talker, and a single-resident lazy interpreter cache for the bucketed units
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.

Using this model

Install dependencies

uv python install
uv sync --frozen

The environment is locked against Python 3.14 for Linux on arm64 (glibc 2.28 or newer); every dependency resolves to a prebuilt wheel, so no compiler is needed.

All cascade units listed in config.yaml must be present; the example resolves them by name from the downloaded directory.

Run the example

uv run example.py

Note: The Python/uv example runs on AWS Graviton (Ubuntu arm64) to confirm runtime compatibility only, and is intended as a guideline for building an equivalent run script on Mobile CPU systems.

With no arguments the example synthesizes the sentence that produced the committed sample_output.wav, using sample_input.wav as the reference speaker, and writes output.wav in the current directory. Pass --text, --ref-audio and --out to change any of the three.

Expected input

Property	Value
Input shape	text tokenized to input_ids of shape [1, T]; reference audio as a mono waveform [N]
Input type	int64 input_ids; float32 reference audio
Input range	reference audio in [-1.0, 1.0] at 24000 Hz
Preprocessing	Tokenize the UTF-8 text with the Qwen3-TTS processor, then resample the reference clip to 24000 Hz; no further normalization

Expected output

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 24000 Hz
Postprocessing	Autoregressive talker decode of the EOS-terminated codec token stream, sub-talker residual-codebook decode over 16 code groups with KV-length-bucketed dispatch, then codec decoding to a waveform with length-bucketed dispatch

Intended use

This model is intended for developers evaluating text-to-speech and voice-cloning workloads on Arm-based platforms. It is suitable as a reference implementation for benchmarking, prototyping, and integration exploration.

Limitations

• Performance depends on the target device, runtime version, backend/delegate support, memory configuration, and system load.
• Accuracy was evaluated on the English subset of tts_multilingual and may not generalize to all domains or languages.
• This release preserves the original model's intended task and behavior, but users should validate it for their own application, data, and deployment environment.
• This repository is not a replacement for the original model documentation.

Additional notes

• No per-layer skip list is used. The LiteRT quantizer does not expose a prepare-time filter, so precision is assigned per whole cascade unit. The speaker encoder and sub-talker backbone are deliberately kept at FP32 to preserve speaker similarity and prefill stability.
• Running the example needs roughly 14 GB of RAM. Peak resident set size was 13.5 GB for a short utterance and 14.5 GB for a longer one.

About this version

Original Model: Qwen/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.

Original Model and Documentation

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.

Licenses and Third-Party Terms

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.

Purpose of this Release

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.