TTS-V4 Chinese Text-to-Speech Model: 8GB+ VRAM, File List

TTS-V4 Chinese Text-to-Speech Model: 8GB+ VRAM, File List

At a Glance

Item Value
Repository autotrust/TTS-V4
Publisher guide Alibaba (Qwen): models and licenses
Published 2026-11-16
License other
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 a Chinese text-to-speech model for call centers, “autotrust/TTS-V4," has been released on Hugging Face. Please note that information regarding this model includes unverified details that have not been officially confirmed, and the report is based on a public repository shared by the community.

TTS-V4 is a Text-to-Speech (TTS) model that performs adaptive learning on the base model “Qwen/Qwen3-TTS-12Hz-1.7B-Base" using actual telephone recording data. It is designed for verbatim reading of standardized scripts such as customer support, sales, and debt collection, and is provided as a fully merged BF16 model that does not require additional loading of LoRA or similar adapters. Furthermore, it is reported to support bidirectional streaming, allowing it to generate audio sequentially while receiving text output from a Large Language Model (LLM) in real-time.

Specifications

The main specifications of the model as described in the public documentation are as follows:

  • Base Model: Qwen/Qwen3-TTS-12Hz-1.7B-Base
  • Architecture: Qwen3TTSForConditionalGeneration (non-DiT discrete multi-codebook LM architecture)
  • Format: BF16 (safetensors format, merged model)
  • Supported Language: Chinese (zh)
  • Audio Output: 24 kHz sampling rate (PCM float32; supports 8000 Hz / PCM 16-bit via WebSocket), supports sequential output in 80 ms frame increments
  • Preset Voices: 6 fixed voices (sales_female, sales_male, service_female, service_male, bank_collection_male, bank_collection_female); also supports custom specification using arbitrary reference audio and transcript text
  • Inference Modes:
  • Voice Mimicry Mode: icl (In-Context Learning, uses reference audio + transcript to reproduce natural intonation), xvec (uses speaker vector only for immediate generation start) – Text Input Method: sentence (sentence-level, split by punctuation or max 60 characters), token (sequential generation per input token)
  • Speed Adjustment: Features time-stretching while maintaining pitch via the WSOLA algorithm (speed multiplier: 0.5–2.0, recommended 0.8–1.3)
  • License Conditions: The distribution license is specified as other. Users must verify the base model’s conditions and the included license files (LICENSE.md, LICENSE-QWEN.txt, NOTICE). It is explicitly stated that this is not an official Qwen model.

Performance

The model card in the public repository includes benchmark results using an independent test set (15 calls, 49 sentences total) not used for training or model selection. Evaluation metrics reported include CER (Character Error Rate, lower is better), AUC for a synthetic vs. real voice classifier (closer to 0.5 is closer to real voice), and speaker similarity.

→ Scroll horizontally to see all columns

Model CER Whisper Feature AUC Prosody AUC Speaker Similarity
Qwen Base (Default Sampling) 0.74% 0.945 0.776 0.786
V3 0.89% 0.906 0.754 0.802
V4 Offline (T=0.8) 2.17% 0.867 0.723 0.772
V4 Streaming icl (T=0.8) 1.48% 0.883 0.770 0.776
V4 Streaming xvec (T=0.8) 0.79% 0.948 0.769 0.758

According to the published data, TTS-V4 (offline, temperature 0.8) achieves closer values to 0.5 in Whisper feature AUC (0.867) and prosody AUC (0.723) compared to the base model and the previous V3 generation, suggesting it has acquired more natural timbre and intonation closer to real telephone audio. On the other hand, the CER, which measures character-level transcription errors, degraded from 0.74% in the base model to 2.17% in the offline version; trade-offs such as ambiguous pronunciation of certain characters, inclusion of filler words, or misreading of specific terms have been reported. However, it is stated that the CER is suppressed to 0.79% in the streaming xvec mode.

Additionally, the following latency and ASR error counts were shown using a long business script of approximately 230 characters (sequential input of 30 characters per second, temperature 0.8):

→ Scroll horizontally to see all columns

Combination service_female First Packet bank_collection_female First Packet Total Playback Stutter ASR Errors (180 chars total)
icl + sentence 593 ms 501 ms 0.12–0.13 s 0 / 2
icl + token 3724 ms 2275 ms 0.12 s 0 / 1
xvec + token 210 ms 217 ms 0.38 s 1 / 1
xvec + sentence 487 ms 493 ms 0.12 s 0 / 0

Experimental results show that while the xvec + token combination achieves an extremely fast first-packet latency of approximately 210–217 ms, the cumulative stutter (interruptions) during playback increases slightly to 0.38 seconds. The recommended icl + sentence setting requires about 500–593 ms for the first packet, but keeps playback stutter to around 0.12–0.13 seconds, which is described as balancing voice reproducibility with stable playback.

Strengths and Use Cases

This model is reportedly specialized for the verbatim reading of standardized scripts in call centers, such as customer support, telemarketing, and debt collection. Its strengths include reading without unnecessary colloquial rewriting of input text, maintaining fidelity to the original text while using timbre and prosody close to actual telephone recordings.

It is reported that 6 fixed voices are provided as presets: sales (female/male), customer service (female/male), and debt collection (female/male). It is also possible to generate audio using custom voices by providing arbitrary reference audio and transcript text.

It is strongly intended for use in real-time voice interaction systems linked with LLM streaming output, and the following settings are proposed depending on the objective:

  • icl + sentence (Recommended mode): A method that generates text in sentence-level chunks. By combining it with the icl mode, which deeply mimics the nuances of reference audio, it can reproduce the most natural human-like timbre and intonation. The latency from the start of the LLM response to the first audio output is about 0.5 seconds, offering an excellent balance between conversational naturalness and latency.
  • xvec + token (Ultra-low latency mode): A method that starts generation immediately as each character (token) is input. Combined with the xvec mode, which uses only speaker vectors, it can begin audio output with an extremely low first-packet latency of about 0.2 seconds after receiving the first character. While naturalness and pronunciation accuracy are slightly lower than icl, it is considered suitable for interaction systems where immediate response is the top priority.
  • Time-stretching (speed parameter): Applies WSOLA processing to generated audio, allowing the speech rate to be adjusted from 0.5x to 2.0x (recommended 0.8–1.3x) without changing the pitch. It is reported that CPU load is very light even during streaming, allowing speed changes without increasing first-packet latency.
  • Automatic Text Normalization: Features the ability to automatically convert numbers, currency, dates, and English text into appropriate Chinese readings (e.g., converting “235.60元" to “两百三十五点六零元").

If a user interrupts the speech during synthesis, the generation process can be immediately aborted using the cancel() method or a cancellation command via WebSocket, making it suitable for building interactive voice AI agents.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 1.9B parameters

Your VRAM Quantization File size Est. memory needed
8GB (RTX 4060 / 3060 Ti, etc.) BF16 3.6GB 4.3GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: registered, vLLM: not registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.

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.

License — other: 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 (other). 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-V4, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
adapters/v4/adapter.safetensors 42MB
model.safetensors 3.86GB
speech_tokenizer/model.safetensors 682MB

How to Get It

This model is distributed from the Hugging Face repository autotrust/TTS-V4 and can be downloaded as a fully merged model (BF16) in safetensors format. It is reported that no gated access requiring prior approval is set, and it can be obtained directly. However, since the license type is defined as other, it is recommended to check the LICENSE.md and the base model’s terms of use included in the repository beforehand.

To use it, install the necessary dependencies in a CUDA-compatible PyTorch environment.

pip install -r requirements.txt
python verify_files.py

For offline environments where you want to synthesize the entire text at once, you can use the included synthesize.py script as follows:

python synthesize.py --voice service_female --text "您好,您本期账单金额是235.60元。" --output out.wav

If you want to simulate LLM streaming output and generate audio sequentially, you can execute it by specifying the inference mode (--stream) and speed multiplier (--speed):

python synthesize.py --voice service_female --stream icl --cps 30 \
  --text "您好,很抱歉让您久等了。您的退款预计3个工作日内到账。" --output icl.wav

Alternatively, when calling from a Python program, import the TTSV4 class and use tts.stream() or tts.stream_sentences(). A WebSocket server script (stream_server.py) is also included for operation as a backend server, allowing the construction and use of a streaming audio synthesis API on port 8765 by default.

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.