TTS-V1 Chinese Text-to-Speech Model: File List

Sample outputs are available on the model card.
At a Glance
| Item | Value |
|---|---|
| Repository | autotrust/TTS-V1 |
| Published | 2026-11-13 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
The open-weight model “autotrust/TTS-V1" has been released, capable of synthesizing speech from Chinese text.
This model takes Chinese text as input and generates speech in 24 kHz mono WAV format. It utilizes a shared architecture that integrates FlowLM and Mimi within a single model file, allowing users to switch between four different voice presets (male and female) intended for sales and customer support applications.
Specifications
The specifications listed in the model card and repository are as follows:
- Architecture: Shared model consisting of FlowLM and Mimi (settings defined in
config.yaml) - Output Specifications: 24 kHz sampling rate, mono, WAV format
- Supported Languages: Chinese (the Chinese tokenizer
tokenizer.modelis included in the repository) - Selectable Voice Presets: Four types are provided, with fixed reference recordings stored in the
voices/directory: sales_male(Male voice for sales/marketing) –sales_female(Female voice for sales/marketing) –service_male(Male voice for customer support) –service_female(Female voice for customer support)- Text Preprocessing: Uses normalization scripts for currency, dates, and various numbers (
tn.py) andwetextresources for speech-text normalization - Segment Control: Automatic text segmentation based on punctuation and token length
- Inference Control Options: Includes
--attempts(default 16) to control maximum retries per segment,--seed(e.g., 3000) for random seed, and--threads(e.g., 2) to specify CPU thread count - Distribution Format: safetensors format (
model.safetensors) - Inference Engine: Includes a full set of Pocket TTS inference code adapted for this model (
pocket_tts/) - License and Terms of Use: The model card directs users to check
MODEL_CARD.mdin the repository for information regarding the origin and distribution terms of the weights and reference audio.
Performance
The correspondence between the parameter names and audio samples provided by the publisher is as follows:
| Parameter | Voice | Sample |
|---|---|---|
sales_male |
Sales Male | Listen |
sales_female |
Sales Female | Listen |
service_male |
Service Male | Listen |
service_female |
Service Female | Listen |
According to the model card, the names of each voice are tentative, based on intended use and pitch; it is recommended to listen to the provided samples (WAV files) to confirm if they meet actual business requirements.
Regarding the quality of the generation pipeline, the included inference script (synthesize.py) incorporates an automatic verification mechanism. When text is segmented by punctuation or token length for sequential synthesis, it checks for valid end-of-sentence (EOS) tokens, ensures basic volume levels, and detects unnatural silence within the audio. Segments that fail to meet these criteria are automatically re-sampled (retried). If the maximum number of attempts is reached without success, the design halts the final file write and reports an error.
However, the publisher notes that even with this verification logic, it is not possible to completely eliminate all noise, typos, or misread numbers. Therefore, for production use, human listening or verification via automatic speech recognition (ASR) after generation is required. Additionally, among the four provided voices, the customer support male voice (service_male) is reported to have a tendency toward unstable termination behavior, sometimes requiring multiple attempts to achieve a normal end state.
Strengths and Use Cases
This model is suitable for speech synthesis (TTS) specialized for business scenarios and interactive applications in Chinese. It is intended for use cases such as guidance messages on e-commerce sites or in sales activities, and the construction of automated voice response systems for customer service.
Synthesis examples shown in the model card include phrases like “您好,您的订单金额是235.60元,预计2个工作日到账。" (Hello, your order amount is 235.60 yuan, expected to arrive within 2 business days), demonstrating its suitability for reading notification messages containing order amounts, dates, and processing times.
To support such operations, the repository includes a normalization script (tn.py) that automatically preprocesses numbers, dates, IDs, and symbols into appropriate readings. This allows for smooth conversion to speech even for business text containing a mix of characters and numbers.
Furthermore, this model allows users to select from four voice personas (sales male, sales female, service male, service female) defined by purpose and pitch from a single shared model file. It is possible to build a system that switches between bright-toned sales voices and polite, conversational customer support voices depending on the application. As noted, official voice names are tentative, so it is recommended to verify them against provided samples.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch) — 115M parameters
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 4GB (laptop iGPU / phone class) | F32 | 0.4GB | 0.5GB |
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. The listed files are distributed in F32 (32-bit). Loading in a lower precision (BF16 or quantized) needs less memory, so treat this as the size of the files as distributed, not a minimum. This site has not measured the lower-precision figures. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Our Own Measurements
Measurements We Did Not Take
This model does not state a license, so we cannot confirm permission to use it. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Distributed Files
Weight files published in autotrust/TTS-V1, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
model.safetensors |
459MB |
How to Get It
Model weights and related code are available in the Hugging Face repository autotrust/TTS-V1. The model is distributed as a single model.safetensors file, which contains both FlowLM and Mimi components.
Linux and Python 3.12 are supported. Installation steps and dependency configurations are as follows:
First, install PyTorch according to your environment. If running in a CPU environment, select the CPU version of PyTorch; if using an NVIDIA GPU, install the version of PyTorch corresponding to your graphics card and driver CUDA version.
Next, create a virtual environment within the directory where you cloned or downloaded the repository and install the necessary libraries.
python3.12 -m venv.venv
source.venv/bin/activate
python -m pip install --upgrade pip
# For CUDA environments, install the appropriate version of PyTorch first
python -m pip install -r requirements.txt
The verified dependency configuration is Python 3.12, torch 2.13.0, NumPy 2.3.5, soundfile 0.14.0, and wetext 0.2.1. Note that an internet connection may be required when installing the wetext package and preparing its speech-text normalization resources. The repository includes Pocket TTS inference code (pocket_tts/) tuned for this model.
Speech synthesis is performed from the command line using the included synthesize.py script. The basic command to specify text and generate a WAV file using the sales female voice (sales_female) is as follows:
python synthesize.py --voice sales_female \
--text '您好,您的订单金额是235.60元,预计2个工作日到账。' \
--output result.wav
To synthesize from a text file using the service male voice (service_male), use the --text-file option:
python synthesize.py --voice service_male --text-file text.txt --output result.wav
The following parameters are available for inference:
--device cuda: Uses an NVIDIA GPU for faster inference (defaults to CPU if not specified).--attempts 16: Controls the maximum number of retries for each text segment.--seed 3000: Specifies a random seed to ensure reproducibility.--threads 2: Specifies the number of threads for CPU execution.
If a WAV file with the same name already exists at the output destination, the program is designed to refuse the operation to prevent overwriting. For details on the origin of the model weights and reference audio, as well as detailed distribution terms, please refer to MODEL_CARD.md included in the repository.
Related Articles
- TTS-V5 Chinese TTS Model: 4GB+ VRAM, File List
- TTS-V4 Chinese Text-to-Speech Model: 8GB+ VRAM, File List
- TTS-V3 Phone-Dialogue Specialized TTS Model: 8GB+ VRAM, File List
- Qwen3.8-27B-NVFP4-DGX-SPARK-V2 NVFP4 Quantized 27B Model: 24GB+ VRAM
What to Read Next
- Find models by VRAM (This model runs from the 4GB tier) → Other models that run on a 8GB GPU
- Formats this model is available in → Safetensors format guide and models

