{"kind":"compare","slug":"fish-audio","path":"/compare/fish-audio","title":"Fish Audio Alternative That Runs on CPU | SuperClonica","description":"Fish Audio S2 Pro and Fish Speech compared with SuperClonica: CPU speed, quality, languages, licenses, and when the heavier Fish Audio model is the right pick.","h1":"Fish Audio and Fish Speech vs SuperClonica","intro":"Fish Audio makes some of the most expressive open text-to-speech models, and our own comparison gives its S2 Pro the top quality mark. It is also a large model built for GPUs. SuperClonica goes the other way: a small Supertonic 3 voice file that runs fast on an ordinary CPU, offline.\n\nIf you need a Fish Audio or Fish Speech alternative because you have no GPU or want to ship speech inside an app, read on. If you need maximum expressiveness or Chinese, Fish Audio is probably the better tool.","sections":[{"heading":"Fish Audio, Fish Speech and S2 Pro: what is what","paragraphs":["Fish Audio publishes its open models in a GitHub repository called fish-speech, and earlier releases carried that name: Fish Speech 1.5, for example, covers 13 languages under CC-BY-NC-SA-4.0. Newer models are called Fish Audio S2 and S2 Pro.","S2 Pro, the model in our table, is dual-autoregressive: a 4B-parameter model predicts the main codebook and a 400M-parameter model fills in the other nine. Its model card lists 80+ languages and free-form tags inside the text, such as [whisper in small voice] or [pitch up]. The repository now also lists Fish Audio Drama 3, a newer and larger model our table does not cover yet."]},{"heading":"Fish Audio S2 Pro vs SuperClonica: speed on CPU","paragraphs":["By our estimate for a desktop 8-core CPU, S2 Pro needs 4 to 10 seconds of computing per second of audio, so a 10-second line takes 40–100 seconds. SuperClonica produces it in about 1.0–1.6 seconds.","This is the roughest figure in our table. There is no public CPU benchmark, so we derived it from the s2.cpp GGUF client and the architecture, assuming 30–50 GB/s of effective memory bandwidth. On a GPU things look different: Fish Audio reports a real-time factor of 0.195 and about 100 ms to first audio on a single NVIDIA H200."]},{"heading":"Where Fish Audio is stronger","paragraphs":["Fish Audio S2 Pro wins on several points that matter for expressive speech."],"points":["Quality: 5 out of 5 in our table, against 4.5 for SuperClonica.","Languages: 80+, including Chinese. Supertonic 3 covers 31 and has no Chinese yet.","Cloning on the fly: according to the fish-speech repository, 10–30 seconds of reference audio is enough to reproduce a timbre, with no training step.","Emotion control: free-form tags in the text for whispering, laughter, emphasis or pitch."]},{"heading":"Licenses and commercial use","paragraphs":["The S2 Pro weights are under the Fish Audio Research License: research and non-commercial use are free, commercial use needs a separate license from Fish Audio. The repository's code and other weights fall under the Fish Audio Community License.","Supertonic 3 uses OpenRAIL-M, a royalty-free license with a list of prohibited uses (impersonating people without consent is one) rather than a non-commercial clause. A voice bought on SuperClonica is for your own projects and may not be resold."]},{"heading":"Fish Audio or SuperClonica: how to choose","paragraphs":["SuperClonica leads on CPU speed and timbre stability: we rate it 5 out of 5 on both, against 1 and 4 for S2 Pro. For hours of narration on a laptop, or speech generated on a player's machine, that counts for more than the last half-star of quality."],"points":["Choose Fish Audio S2 Pro if you have a capable GPU, need Chinese or a rarer language, want instant cloning from a short clip, or rely on emotion tags.","Choose SuperClonica if speech has to run on a CPU, offline, at predictable speed, or inside an app you ship to users."]}],"faq":[{"question":"Is there a free alternative to Fish Audio that runs on CPU?","answer":"Supertonic 3 is open-weight, comes with preset voices and needs no GPU. Our catalog adds free and paid voices, and by our estimate SuperClonica runs them 6.4–9.7× faster than real time on a desktop 8-core CPU. It won't match S2 Pro's expressiveness, but it works where S2 Pro is very slow."},{"question":"Can Fish Speech run without a GPU?","answer":"Fish Audio's own figures are for GPUs. A community client, s2.cpp, runs S2 Pro through ggml. Our rough estimate for a desktop CPU is 4–10× slower than real time, so one minute of speech would take 4 to 10 minutes."},{"question":"Can I use Fish Audio S2 Pro commercially?","answer":"Not under the default terms. The model card says research and non-commercial use are free, while commercial use requires a separate license from Fish Audio."},{"question":"Does SuperClonica clone voices the way Fish Audio does?","answer":"No. Fish Audio clones at generation time from a 10–30 second clip. SuperClonica trains a separate voice file from a 5- to 10-second recording, on your own NVIDIA GPU for free or by an author you commission. Training takes longer, but the file then runs fast on any CPU."}],"cta":{"catalog":"Listen to Supertonic 3 voices","install":"Install the SuperClonica engine"},"sources":[{"label":"Fish Audio S2 Pro model card (Hugging Face)","url":"https://huggingface.co/fishaudio/s2-pro"},{"label":"fish-speech repository (GitHub)","url":"https://github.com/fishaudio/fish-speech"},{"label":"Fish Speech 1.5 model card","url":"https://huggingface.co/fishaudio/fish-speech-1.5"},{"label":"s2.cpp: S2 Pro inference in ggml","url":"https://github.com/rodrigomatta/s2.cpp"},{"label":"Supertonic 3 model card","url":"https://huggingface.co/supertone-oss-archive/supertonic-3"}],"published":"2026-10-17","updated":"2026-10-23","pages":[{"slug":"text-to-speech","path":"/text-to-speech","h1":"Text-to-speech that runs on your own computer"},{"slug":"supertonic","path":"/compare/supertonic","h1":"Supertonic vs SuperClonica: same model, faster and cleaner"},{"slug":"fish-audio","path":"/compare/fish-audio","h1":"Fish Audio and Fish Speech vs SuperClonica"},{"slug":"omnivoice","path":"/compare/omnivoice","h1":"OmniVoice vs SuperClonica: more languages or more speed"},{"slug":"qwen3-tts","path":"/compare/qwen3-tts","h1":"Qwen3-TTS vs SuperClonica for local text-to-speech"},{"slug":"voxcpm","path":"/compare/voxcpm","h1":"VoxCPM 2 vs SuperClonica: studio sound or CPU speed"},{"slug":"f5-tts","path":"/compare/f5-tts","h1":"F5-TTS vs SuperClonica: zero-shot cloning or fast CPU speech"},{"slug":"elevenlabs","path":"/compare/elevenlabs","h1":"An offline alternative to ElevenLabs"},{"slug":"kokoro","path":"/compare/kokoro","h1":"Kokoro TTS vs SuperClonica: two small models for local speech"},{"slug":"piper","path":"/compare/piper","h1":"Piper TTS vs SuperClonica: tiny devices or custom voices"},{"slug":"xtts","path":"/compare/xtts","h1":"Coqui XTTS-v2 vs SuperClonica: a Coqui TTS alternative"},{"slug":"chatterbox","path":"/compare/chatterbox","h1":"Chatterbox TTS vs SuperClonica: instant cloning or a trained voice file"}],"related":[{"slug":"offline-text-to-speech","path":"/guides/offline-text-to-speech","h1":"How to convert text to speech offline on your computer"},{"slug":"voice-cloning","path":"/guides/voice-cloning","h1":"Voice cloning: how to get a voice that sounds like a recording"},{"slug":"best-free-tts-models","path":"/guides/best-free-tts-models","h1":"The best free text-to-speech models for local use in 2026"},{"slug":"text-to-speech-for-videos","path":"/guides/text-to-speech-for-videos","h1":"Text to speech for videos, audiobooks, games and apps"},{"slug":"supertonic-3-voices","path":"/guides/supertonic-3-voices","h1":"Supertonic 3 voices: where to get custom voices and how to use them"},{"slug":"python-text-to-speech","path":"/guides/python-text-to-speech","h1":"Text to speech in Python: offline, free and with natural voices"},{"slug":"text-to-speech-api-docker","path":"/guides/text-to-speech-api-docker","h1":"Self-hosted text-to-speech API in Docker"}]}