Overview
GLM-TTS is a text-to-speech system for generating expressive speech from text. Zhipu AI announced it on December 10, 2025, and released model weights for local deployment. (zhipuai.cn)
Capabilities and uses
- Supports zero-shot voice cloning using approximately 3–10 seconds of reference audio, emotional expression, pronunciation control and streaming inference. The model card describes optimization for Chinese, English and mixed-language text. (zhipuai.cn)
- Uses an autoregressive text-to-token stage followed by flow-matching acoustic generation and a vocoder. Multi-reward reinforcement learning targets pronunciation, speaker similarity and expressiveness. (zhipuai.cn)
- Potential applications include customer service, audiobooks, interactive assistants, education and spoken office content. (docs.bigmodel.cn)
Limitations
Speech quality, pronunciation and voice similarity should be evaluated for the intended text, language and speaker; benchmark results do not guarantee performance in every setting. Voice cloning requires suitable reference audio and appropriate authorization. The cited materials do not establish a complete supported-language list or a universal maximum input length. Features and limits may differ between the open model and hosted API. (huggingface.co)
Pricing
Checked November 17, 2026. The official mainland China BigModel price for GLM-TTS speech generation is CNY 2 per 10,000 characters. The price table lists a character-based charge rather than separate input-token and output-token prices; it does not specify a separate audio-output charge. The same table lists GLM-TTS-Clone, a distinct voice-cloning service, at CNY 6 per call. Actual charges depend on the applicable service and account billing. (docs.bigmodel.cn)
The official international Z.AI pricing page was also checked; a GLM-TTS USD price could not be verified there. (docs.z.ai)