TTS-V5 Chinese TTS Model: 4GB+ VRAM, File List

TTS-V5 Chinese TTS Model: 4GB+ VRAM, File List

At a Glance

Item Value
Repository autotrust/TTS-V5
Published 2026-11-17
License mixed-components
Formats safetensors / ONNX
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

[Unverified] This news has not been officially confirmed. The following is based on the description in the Hugging Face model card (autotrust/TTS-V5).

autotrust/TTS-V5 is an experimental text-to-speech model reported to be a fine-tuned version of FunAudioLLM/Fun-CosyVoice3-0.5B-2512 specialized for Chinese call center domains. It supports the zero_shot method for cloning voices from reference audio. According to the card, the repository includes merged LoRA weights, meaning there is no need to separately download LoRA or base models during inference.

Specifications

  • Base Model: FunAudioLLM/Fun-CosyVoice3-0.5B-2512 (language model initialized from llm.rl.pt)
  • Output: 24 kHz mono WAV. The tagged language is zh
  • Bundled Voices: 6 voice profiles (sales_female, sales_male, service_female, service_male, bank_collection_female, bank_collection_male)
  • Custom Voices: Specify ref_audio and ref_text. Reference audio of 3 to 12 seconds is recommended (maximum 30 seconds, 16 kHz or higher)
  • Generation Modes:
  • Offline synthesis – Audio streaming (--stream) – Bidirectional streaming with incremental text input (simulated with --cps, where cps is not speech rate) – Tone and speech rate instructions via --instruct (not a strict multiplier) – Speech rate post-processing via --speed (WSOLA, range 0.5–2.0, recommended 0.8–1.3). It is noted that audio degradation may become more noticeable the further the value is from 1.0
  • Not full-duplex. According to the card, there are no APIs for real-time audio input, ASR/VAD, interruption detection, playback interruption, or generation cancellation
  • Training: LoRA applied to the last 12 layers of the LM and the audio output head (rank 8 / alpha 16, 2,260,808 trainable parameters, 800 steps, batch size 4, peak learning rate 8e-5). LoRA also applied to the last 8 layers of Flow (rank 8 / alpha 16, 1,376,256 parameters, 400 steps, peak learning rate 3e-5). Vocoder and token extractor remain unchanged
  • License: Listed in documentation as mixed-components (tag license:other). The documentation contains no details on terms, and commercial usability is unconfirmed. The base model is apache-2.0
  • Environment: Transformers must be pinned to 4.51.3. According to the card, version 4.57.3 caused misreadings and early stopping during cache decryption

Performance and Quality

The model card provides no comparison table with other models. The figures provided by the publisher are limited to the following validation metrics during training, and the publisher itself notes that these are “not listening evaluations."

  • LM CE: 4.378 → 4.122 (fixed 48 segments)
  • Flow loss: 0.945 → 0.896 (fixed 24 segments)

Additionally, the card reports the following actual measurements:

  • In a long-form speech test (V1) with incremental text input at 30 characters per second, the time to the first packet was approximately 2.37 seconds. Audio reportedly began playing before the text input concluded
  • ASR typo count was 0/180 for both incremental output and 0.9x speed output

The card explicitly states that this merely demonstrates the operation of bidirectional text and audio streaming, and is not proof of “listen-and-speak" capability. The official “150 ms" latency is also not guaranteed under arbitrary conditions.

Base Model Performance (Reference)

Below are the publisher’s figures from the original Fun-CosyVoice3-0.5B-2512 card, which do not represent the performance of this model. Lower CER and WER indicate better performance, while higher Speaker Similarity is better.

→ Scroll horizontally to see all columns

Model test-zh CER (%) ↓ test-zh Speaker Similarity (%) ↑ test-en WER (%) ↓ test-en Speaker Similarity (%) ↑ test-hard CER (%) ↓ test-hard Speaker Similarity (%) ↑
CosyVoice2 1.45 75.7 2.57 65.9 6.83 72.4
Fun-CosyVoice3-0.5B-2512 1.21 78.0 2.24 71.8 6.71 75.8
Fun-CosyVoice3-0.5B-2512_RL 0.81 77.4 1.68 69.5 5.44 75.0

The base model outperforms CosyVoice2 in both CER and speaker similarity. While the RL version achieves the lowest CER/WER, the speaker similarity for the vanilla version is higher across all three test sets (test-zh, test-en, and test-hard).

Strengths and Use Cases

Although unverified information, the model card indicates this model is an experimental version targeting Chinese call center audio. The 6 bundled voices are categorized by use case:

  • Sales: sales_female, sales_male
  • Customer Service: service_female, service_male
  • Bank Collection: bank_collection_female, bank_collection_male

While the original Fun-CosyVoice3-0.5B-2512 is a general zero-shot TTS supporting 9 languages and over 18 Chinese dialects, this model is adjusted with LoRA focusing specifically on Chinese for call centers. Example sentences in the card consist of standardized response scripts, such as refund arrival notifications and amount confirmations.

Anticipated use cases include:

  • Audio output for response bots that sequentially read text generated by an LLM. The card’s tts.stream() can directly accept incremental LLM outputs. Internally, it normalizes at punctuation positions to avoid misreading number sequences that have not yet arrived.
  • Custom voice cloning created from internal call materials. When using telephone audio for reference speech, it is recommended to select clear segments that do not overlap with the interlocutor’s speech. It is also noted that resampling will not restore the original telephone line bandwidth.
  • Response audio with varied speech rates or tones. --instruct is a method where the model interprets tone and speech rate, while --speed is post-processing that precisely specifies a multiplier.

Constraints are also explicitly stated in the card:

  • It is not a telephone-dialogue-level full-duplex system, and interruption handling is not implemented.
  • The Python wrapper performs sequential generation and serves as an experimental entry point sharing the model.
  • The V4 WebSocket session protocol is also not implemented.
  • Behavior in other languages is not documented in the materials.

Furthermore, the figures provided in the card are limited to training validation metrics and select actual measurements. Perceptual quality must be verified independently using the bundled listening pack.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch)

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) Original precision(CosyVoice-BlankEN) 0.9GB 1.1GB

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. 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 / ONNX.

License — mixed-components: 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 (mixed-components). 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-V5, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
campplus.onnx 28MB
CosyVoice-BlankEN/model.safetensors 988MB
flow.pt 1.33GB
hift.pt 83MB
llm.pt 2.57GB
speech_tokenizer_v3.onnx 969MB
training/flow_0400.pt 6MB
training/llm_0800.pt 9MB

How to Get It

Distributed formats are safetensors and ONNX. According to the card, the repository root contains pre-merged llm.pt and flow.pt, original vocoder and token extractor, text model assets, 6 voice reference samples, and inference code. There is no need to separately download LoRA or base models. gated is false, allowing access without license agreement. However, the license is listed as mixed-components, and detailed conditions are not provided in the documentation. Please check MODEL_CARD.md or similar files in the repository prior to commercial use.

Download commands are as specified in the card:

pip install "huggingface-hub==0.36.2"
python -c "from huggingface_hub import snapshot_download; snapshot_download('autotrust/TTS-V5', local_dir='TTS-V5')"
cd TTS-V5

Next, in a Python environment with CUDA-enabled PyTorch and torchaudio installed, install dependencies and verify the files:

pip install -r requirements.txt
python verify_files.py

Example for offline synthesis:

python synthesize.py --voice service_female \
  --text "您好,您的退款预计3个工作日内到账,请注意查收。" --output out.wav

Appending --stream enables audio streaming. Combining --stream --cps 30 simulates an incremental text input rate of 30 characters per second. Call it from Python as follows:

from inference import TTSV5
tts = TTSV5()
pcm = tts.synthesize("您好,请您核对一下金额。", voice="service_female")

When using streaming, ensure enqueue_audio() returns promptly. Blocking until the current chunk’s playback finishes may impact the generation schedule of subsequent chunks, as stated in the card. Regarding the environment, as noted in the first section, pin Transformers to 4.51.3.

The repository also includes a 6-voice comparison listening pack TTS-V5-listening.zip. Unzip and open listening/index.html to listen to generation examples. Detailed results can be found in MODEL_CARD.md, RESULTS.md, and evaluation/.

Information regarding compatible tools such as ComfyUI or A1111 is not mentioned in the documentation.

Related Articles

What to Read Next

Sources

This article contains unverified information. We will append an update note once it is confirmed by a primary source.