Fish Audio: cloned voices, a free model, word highlighting

English · 简体中文

Fish Audio is a voice-cloning service: every voice on it was made from a short recording — by its users, or by you — and its current model, S2.1 Pro, speaks 83 languages. The plugin's Fish Audio section reaches it with one API key, and Read Aloud highlights word by word with these voices, as it does with Azure's and Kokoro's. This page: the key, what is free and what costs, and how a voice gets into the player.

The key

1. Sign up at fish.audio, click Open App, and choose Developer at the top of the left sidebar.
2. API Keys → Create API Key (or go straight to fish.audio/app/api-keys): name it, leave Expiration on Never — reading stops the day a key expires — and click Create. Copy the key right away: the list shows it masked afterwards.
3. In Zotero, Edit → Settings → OpenReader → Fish Audio: paste it into API key, click Test connection, then Enable.

Free or paid

The block's Use only the free model switch is on when you start.

• On — every sentence is spoken by the free S2.1 Pro. It needs no balance on the account. Its terms, in Fish Audio's words: no cap on the amount of text, no guarantee of speed, and the text you send may be kept to improve their models. It was announced as a limited-time offer and has been extended more than once — as of October 2026, through November 30, 2026; it may end.
• Off — the paid S2.1 Pro: the same model, $15 per million bytes of text, about 180,000 English words; a Chinese character is three bytes. It is billed to your API credit, the Balance under Developer → API Billing, filled with Top Up. It is separate from the website's own plan and credits under Creative → Billing. With no API credit, Test connection says Insufficient API credit and nothing plays until you top up or switch the free model back on.

Voices

Voices appear under Fish Audio in the player's first dropdown, by name. Only Official voices starts on; Your voices and Manual voices start off. The three source switches work independently.

• Official voices. Voices published by Fish Official are listed automatically. Community voices are added only when you enter their Model IDs yourself.
• Your voices. Voices created by your Fish Audio account are listed without pasting their IDs: those on the My Voices tab of the website's My Voices page. A voice you only bookmarked or liked is not; add it by its Model ID.
• Manual voices. Offers the voices entered in Voices (Model IDs). Turning this switch off hides them without erasing the saved IDs.
• Voices (Model IDs). Fish Audio calls each voice's identifier its Model ID. Click Model IDs in the field label to open the Fish Audio discovery page. Open a voice's page, click ⋯ → Copy Model Id, and paste it here; the ID is also the last part of that page's address, fish.audio/app/m/<id>. Separate several with commas or spaces. The ? beside the field points to that page. Existing pasted voice links still work. To edit the field, Disable the cloud block first, then Enable it again.
• Changing a source. Disable Fish Audio first to unlock its source choices, then Enable it after making changes. If a player is open, disabling asks to stop reading first. A duplicate voice appears only once and stays listed as long as an enabled source supplies it.
• Language. Voices are grouped by their language, or under Multiple languages when they list several. English voices with a published regional accent are grouped separately, such as English (United States), English (United Kingdom), and English (India). English voices without a region stay under English.
• Default. The model's own voice is included with Your voices, even if your account has no custom voices. Turn Your voices off to hide it. With all three sources off, no Fish Audio voices are available; Test connection still works. If a saved voice becomes unavailable, choose another voice before reading.
• Updating the list. Enable and Test connection reload the Fish Audio voice list. Repeated openings reuse the loaded list until Zotero restarts. A failed list reload keeps the last successful voices with a warning; changing the API key does not reuse another account's list.

Test connection counts the listed voices and checks that the model can speak. A manually entered voice whose page has gone is listed as (not found) until you remove its ID.

Good to know

• A sentence takes about two seconds to arrive, the first one of a session a little longer; the player fetches ahead, so reading is continuous.
• A very long passage — a paragraph without a period, a table read as one sentence — takes proportionally longer: a minute and a half of speech arrives in about forty seconds.
• What is sent is the sentence being read, and the id of the voice.
• Speed is the player's own slider, as with every voice; nothing is regenerated.