GLM-TTS

GLM-TTS is Z.ai's text-to-speech system, published on 10 December 2025 alongside a technical report. It is built in two stages: a language model of the Llama architecture converts written text into speech tokens, and a flow-matching model turns those tokens into mel-spectrograms that a vocoder renders as sound. The language stage holds about 1.55 billion parameters and ships in float32, so the download runs to 6.2 GB before the flow, vocoder and tokenizer modules are counted. The headline feature is zero-shot voice cloning: three to ten seconds of reference audio are enough to reproduce a speaker, with no fine-tuning. On top of that sits the part Z.ai treats as its research contribution — the model was aligned with reinforcement learning (GRPO) against four separate rewards: speaker similarity, character error rate, emotion, and laughter. The stated aim was to fix the flat delivery typical of synthesised speech. The vendor's own comparison table is worth reading closely, because it does not say what a marketing table usually says. GLM-TTS with reinforcement learning scores the lowest character error rate in the field at 0.89, ahead of the closed Seed-TTS (1.12) and of the open CosyVoice2 (1.38) and F5-TTS (1.53). But on speaker similarity — how much the cloned voice actually sounds like the person — the closed model still wins at 79.6 against 76.4. Read plainly: this model reads a script more accurately than anything it was measured against, and still imitates a voice slightly less convincingly than the commercial system it was measured against. Two practical notes belong in any decision to use it. First, the licence is not clean: the model card declares MIT, the GitHub repository carries an Apache 2.0 file, and no licence document accompanies the weights themselves — a discrepancy anyone shipping synthetic speech commercially should resolve with the vendor rather than assume. Second, this model is not for sale. Z.ai's API price list covers speech recognition and nothing else; synthesis exists only as a download and a demo site. That is the reverse of the company's transcription line, where the paid hosted model carries a 30-second cap per call and the free downloadable sibling does not.

▸Videos

No videos yet.

▸Reviews