TTS-V3 Phone-Dialogue Specialized TTS Model: 8GB+ VRAM, File List

Sample outputs are available on the model card.
At a Glance
| Item | Value |
|---|---|
| Repository | autotrust/TTS-V3 |
| Publisher guide | Alibaba (Qwen): models and licenses |
| Published | 2026-11-16 |
| License | component-specific-terms |
| Formats | safetensors |
| Source type | Unverified (not confirmed by a primary source) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
It is reported that “autotrust/TTS-V3", a model for voice cloning and conversational Text-to-Speech, has been released on Hugging Face. Please note, however, that official and definitive information as well as corroboration regarding this model have not yet been obtained, making it unverified information.
This model is reported to output 24 kHz monophonic WAV audio from input text and reference audio data. It is based on the open-weight model “Qwen/Qwen3-TTS-12Hz-1.7B-Base" developed by Alibaba, and was reportedly developed by applying 400 steps of LoRA adaptation using actual telephone conversation data.
It supports a total of 6 types of reference audio across 3 use cases (sales, customer service, and bank collection/payment arrangement) for both male and female voices, aiming to provide stable speech synthesis and voice cloning in Chinese dialogue scenarios.
Specifications
According to the published model card and related materials, the main specifications and conditions of autotrust/TTS-V3 are as follows:
- Base Model: Qwen/Qwen3-TTS-12Hz-1.7B-Base
- Model Structure / Architecture: Qwen3TTSForConditionalGeneration (base model)
- Trained Parameters (Additional): 1,245,184 FP32 parameters (LoRA training increment)
- Output Audio Specifications: 24 kHz monophonic (WAV format)
- Target Languages and Domains: Chinese (zh), telephone conversations (sales, customer service, bank collection)
- Default Configurable Parameters:
- Maximum audio code generation tokens (
--max-new-tokens): 512 – Seed value (--seed): 5000 – Device to use (--device): cuda:0 - Usage Conditions and License:
license:other(Complies with component-specific terms of use. Must follow the respective license conditions described in LICENSE.md and LICENSE-QWEN.txt in the model repository)
The repository provides about 5 MB of trained LoRA adapters and loading code. Actual inference requires downloading the approximately 4.5 GB Qwen base model and audio tokenizer.
Performance
Controlled evaluation experiment results published in the model card are as follows. The evaluation set used data from a total of 72 initial generations (4 groups totaling 288 items) using 6 types of reference audio × 12 independent Chinese speech scripts (short responses, questions, amounts, dates, long texts, etc.).
TTS-V3 Controlled Evaluation Results
→ Scroll horizontally to see all columns
| 模型 | Qwen ASR CER↓ | Whisper CER↓ | 声纹余弦中位数↑(同电话带宽) | 观察到EOS |
|---|---|---|---|---|
| Pocket V1 | 2.73% | 4.95% | 0.546 | 65/72 |
| Pocket保守微调版 | 3.15% | 4.70% | 0.553 | 66/72 |
| Qwen 1.7B原底座 | 4.59% | 12.23% | 0.703 | 72/72 |
| TTS-V3 | 2.12% | 3.30% | 0.707 | 72/72 |
*CER (Character Error Rate) is an index indicating the character-level error rate of transcriptions in speech recognition, where lower values mean better performance.
In the published benchmark results, TTS-V3 recorded 2.12% for Qwen ASR CER and 3.30% for Whisper CER, which is reported to significantly reduce the character error rate compared to the original Qwen 1.7B base model (4.59% and 12.23%) and the comparison Pocket versions. It also showed a median speaker voiceprint cosine similarity of 0.707 in the same telephone bandwidth, exceeding the base model’s 0.703 and demonstrating stable performance.
On the other hand, the model card explanation explicitly states that the reduction in CER and improvement in voiceprint similarity are mainly due to the stabilization of content expression, and do not immediately guarantee MOS (Mean Opinion Score) evaluation indicating naturalness or “human-like naturalness." In comparison with real data, while some of the pitch range and energy dynamics approached real recordings, issues were also noted, such as a lack of consistent improvement in subtle pauses, speech rate, and phrase-ending intonation.
As for the performance evaluation of the base model “Qwen/Qwen3-TTS-12Hz-1.7B-Base" itself, the published content consistency (Word Error Rate: WER↓) scores on the SEED-TTS test set recorded 0.77 for Chinese (test-zh) and 1.24 for English (test-en), showing that it possesses high content consistency compared to major speech synthesis models.
Strengths and Use Cases
autotrust/TTS-V3 excels at conversational speech synthesis and voice cloning specialized for Chinese telephone dialogue scenarios and customer solutions. Because training was conducted using actual telephone dialogue audio data, it is said to be capable of handling practical conversational nuances in call centers and business automation.
The use cases and preset reference audio profiles specifically mentioned in the model card are as follows:
- Sales: Male voice (
sales_male) and female voice (sales_female) - Customer Service: Male voice (
service_male) and female voice (service_female) - Bank Collection: Male voice (
bank_collection_male) and female voice (bank_collection_female)
This repository includes these 6 types of 24 kHz human reference audio data, and by specifying the profile name (--voice) when running the inference script, you can generate audio with the voice timbre and atmosphere of the corresponding speaker.
Additionally, a voice cloning function is supported to reproduce any person’s voice by providing user-prepared short reference audio (--ref-audio) and its accurate transcription text (--ref-text) as inputs. This cloning process generates speech utilizing the voice quality, speaking tempo, and expression style contained in the reference audio. However, it is explained that it does not completely replicate the pitch (F0) curve on a frame-by-frame basis. Also, it should be noted that it is a cloning method dependent on the base model (Base) given the reference audio as a condition, rather than a style controlled by natural language emotion instruction text (such as CustomVoice or VoiceDesign).
As a practical consideration, a normalization module (text_normalization.py) is built-in to automatically convert arithmetic numbers included in the input text, such as monetary amounts, dates, account numbers, and phone numbers, into reading formats actually used for recitation. This makes it possible to adjust and process raw text output from business databases or dialogue systems into natural reading expressions.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 1.9B parameters (taken from the base model Qwen/Qwen3-TTS-12Hz-1.7B-Base)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 8GB (RTX 4060 / 3060 Ti, etc.) | BF16 | 3.6GB | 4.3GB |
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. This release is an adapter (LoRA etc.); the table shows what the base model Qwen/Qwen3-TTS-12Hz-1.7B-Base needs. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Can You Run It Locally?
The publisher distributes this model as safetensors.
License — component-specific-terms: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.
Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.
Our Own Measurements
Measurements We Did Not Take
We have not confirmed that this site meets the commercial-use terms of this model’s license (component-specific-terms). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Recent Models in the Same Size Class
Models with up to 4B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.
→ Scroll horizontally to see all columns
| Model | Parameters | Smallest VRAM tier | License | Our article |
|---|---|---|---|---|
| ggml-org/Soprano-1.1-80M-GGUF | 80M | 4GB | apache-2.0 | Soprano-1.1-80M-GGUF Lightweight English TTS Model: File List (2026-11-17) |
| phasefield-audio/Irodori-TTS-v4.1-Anime | 766M | 4GB | mit | Irodori-TTS-v4.1-Anime: Our Generated Audio, File List (2026-09-07) |
Distributed Files
Weight files published in autotrust/TTS-V3, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
model/adapter.safetensors |
5MB |
How to Get It
The weight data and complete set of inference scripts for autotrust/TTS-V3 can be obtained from the Hugging Face repository. Since prior account agreement procedures (gated model designation) are not set for using the model, anyone can access and acquire it.
Set up the necessary libraries and acquire the repository by running the following commands:
python -m pip install 'huggingface_hub>=0.34,<1'
hf download autotrust/TTS-V3 --local-dir TTS-V3
python -m pip install -r TTS-V3/requirements.txt
As for the operating environment, Linux OS, Python 3.11, and an NVIDIA CUDA environment are recommended. PyTorch 2.13.0, BF16 precision, and SDPA (Scaled Dot-Product Attention) were used in the developer’s actual measurements. It is reported that systematic verification has not been conducted for CPU environments or other platforms.
When performing speech synthesis processing, use the included synthesize.py script. An example command for synthesizing using a preset dialogue voice profile is as follows:
python TTS-V3/synthesize.py \
--voice bank_collection_female \
--text '您好,想跟您核实一下这期账单的还款安排。' \
--output bank_v3.wav
To clone an arbitrary voice using a self-prepared reference audio file, specify the --ref-audio and --ref-text options and run as follows:
python TTS-V3/synthesize.py \
--ref-audio reference.wav --ref-text '参考録音の実際の転写テキスト。' \
--text '合成したい新しいテキスト。' \
--output clone.wav
Note that upon the first execution of the script, the base model “Qwen/Qwen3-TTS-12Hz-1.7B-Base" (fixed revision fd4b254389122332181a7c3db7f27e918eec64e3) and related audio tokenizers are automatically downloaded via Hugging Face’s standard cache specification. If you have already acquired the base model on your local disk, you can also run it by directly specifying the local model directory by adding the --model /path/to/Qwen3-TTS-12Hz-1.7B-Base parameter.
Related Articles
- TTS-V4 Chinese Text-to-Speech Model: 8GB+ VRAM, File List
- TTS-V5 Chinese TTS Model: 4GB+ VRAM, File List
- TTS-V1 Chinese Text-to-Speech Model: File List
- Qwen3.8-27B-NVFP4-DGX-SPARK-V2 NVFP4 Quantized 27B Model: 24GB+ VRAM
What to Read Next
- Formats this model is available in → Safetensors format guide and models
- Learn about the publisher → Alibaba (Qwen): models, licenses and articles
- Other models for the same task → Other audio models
- How to read CER, WER → Benchmark glossary
Sources
This article contains unverified information. We will append an update note once it is confirmed by a primary source.

