Miso TTS 8B is an open-weights text-to-speech model from Miso Labs, released in May 2026 and now sitting at roughly 3,200 GitHub stars. It is a deliberate outlier in a field that has spent the past year racing toward smaller models: where PocketTTS fits in 100M parameters and MOSS-TTS-Nano targets edge devices, Miso TTS commits 8.2 billion parameters to the problem of conversational speech. The trade is explicit — this is not a model you run on a laptop, and the project says so plainly. ## Architecture The model is a text-to-dialogue RVQ Transformer, an architecture the repository credits as inspired by Sesame's CSM design. Two transformers work in sequence. A large Llama-8B-style backbone consumes interleaved text and audio-frame embeddings; a much smaller 300M decoder then autoregressively predicts higher-order audio codebooks within each frame. Audio is tokenized through Mimi, with 32 codebooks and a 2,051-token audio vocabulary against a 128,256-token text vocabulary, at a maximum sequence length of 2,048. The interleaving is the architecturally interesting part. Because the backbone accepts text and audio tokens in the same stream, it conditions generation on conversation history rather than treating each utterance in isolation. That is what the "text-to-dialogue" framing is pointing at — the model is built to produce speech that sounds like a turn in a conversation, carrying prosody and emphasis forward, rather than a sequence of independently synthesized sentences. ## Usage and Voice Cloning The API is compact. A `load_miso_8b` call pulls weights from the `MisoLabs/MisoTTS` Hugging Face repository, and `generate` takes text, a speaker index, an optional context list, and a maximum audio length. Zero-shot voice cloning works by populating that context with `Segment` objects pairing prompt audio with its transcript — no fine-tuning, no separate speaker-embedding step. Setup is a `uv sync` or a pip install into a Python 3.10 environment, and the included `run_misotts.py` produces a sample conversation on first run. Generated audio is watermarked by default through Sony's SilentCipher, downloaded automatically alongside the model. The repository is direct about deployment expectations here: anyone embedding the model in an application is told to use their own private watermark key and keep it secret, and the safety section explicitly warns against impersonation and deceptive audio. ## Hardware Reality The README is unusually candid about system requirements, which is worth noting because this is where most 8B TTS deployments will fail. At bfloat16 the weights are roughly 16 GB and the recommended VRAM is 24 GB — an RTX 3090, 4090, A5000, or L4. Float32 pushes that to ~33 GB of weights and 40 GB or more of VRAM. CPU inference runs but is slow, needing around 20 GB of RAM at bf16. First run downloads 30-40 GB total, covering the checkpoint, the Mimi codec, the SilentCipher watermarker, and the Llama 3.2 tokenizer. The project also preemptively corrects a misconception: public discussion of 110 ms latency refers to Miso Labs' hosted production API on H100-class hardware, not to the unoptimized local inference path in this repository. That kind of disclosure is rare and welcome, but it is also a clear signal — the open release is the model, not a tuned serving stack. ## Pros and Cons The case for Miso TTS is quality at the top of the parameter curve. An 8B backbone with dialogue-aware conditioning targets a different quality ceiling than distilled sub-1B models, and the context-based cloning path is genuinely simple to use. Weights are public on Hugging Face, the inference code is short and readable, and the documentation's honesty about hardware and latency saves real evaluation time. The constraints are equally clear. English only, with no stated multilingual roadmap — a hard stop for many use cases when OmniVoice covers 646 languages and VoxCPM2 covers 30. The 24 GB VRAM floor rules out consumer deployment and most edge scenarios. The license is a Modified MIT variant from Kamino, Inc.: permissive in substance, but it adds an attribution requirement obliging products above 50 million monthly active users or 10 million USD in monthly revenue to display "Miso Labs" in their interface. That is a light obligation, but it is a non-standard term, and legal teams at large companies will need to read it rather than assume MIT. The repository has also been quiet since early June 2026, and the inference path ships unoptimized — no quantization, no GGUF, no batching guidance. ## Outlook Open TTS has bifurcated. One branch chases efficiency, and the results have been striking — sub-100M models generating hours of audio in minutes. The other bets that conversational naturalness needs scale, and Miso TTS is a clean expression of that bet. Its practical future likely runs through the quantization ecosystem: an 8B TTS model at Q8 in a native runtime becomes far more deployable than a 33 GB float32 checkpoint, and projects like audio.cpp have absorbed comparable models quickly. Whether that happens depends partly on whether Miso Labs resumes active development. ## Conclusion Miso TTS 8B suits teams building English conversational voice products with GPU budget to spare, and researchers investigating whether dialogue-context conditioning at scale improves prosody. Anyone needing multilingual coverage, real-time latency on modest hardware, or a fully standard license should look elsewhere — the smaller open TTS models have made that end of the market crowded and capable.