Fish Audio Drama 3 API

Pricing: per UTF-8 byte

Like S2.1 Pro, Drama 3 is billed on the UTF-8 byte length of the input text, not on characters. A Latin letter is one byte; Cyrillic and accented letters are two; Chinese, Japanese and Korean characters are three. Bracketed directions and tags such as <whisper> are part of the text and are billed too.

The special price ends at 2026-11-12 00:00 UTC; billing returns to $0.065 automatically, with no change needed on your side. Texts under 100 bytes are charged as 100 bytes. Failed requests are not charged.

Fish Audio licensed voices (licensed: true in the voice library) are billed at 2.5 times the rate: $0.1625 per 1,000 bytes, or $0.0625 during the special. Community voices and your own clones are billed at the normal rate. If any voice in a multi-speaker request is licensed, the whole request is billed at the licensed rate.

Writing directions

Put a bracketed note before a line to set its emotion, intent or pacing, and the whole line is delivered that way. Directions can be written in English or in your own language; they do not need to match the language of the line.

[barely holding it together, forcing a smile] I'm fine.
[强忍着、挤出笑容] 我没事。

To change only a few words, wrap them in a paired tag. Supported tags:

Really. <whisper>Go on without me.</whisper>

Multi-speaker dialogue

• Pass reference_id as an array of voice ids, up to 16.
• Start each turn with <|speaker:N|>. N counts from 0 and matches the array index: <|speaker:0|> uses the first voice, <|speaker:1|> the second.
• Voice ids can come from the Fish voice library or from voices you cloned at POST /v1/audio/voices/fish.
• With one voice, a plain string still works. Only fish-drama-3 accepts an array; fish-s2.1-pro returns a 400.

Pronunciation and language switching

When a name, brand or term comes out wrong, pin its pronunciation with a phoneme tag. English uses CMU Arpabet (the digit after a vowel marks stress: 1 primary, 0 unstressed), Chinese uses pinyin and Japanese uses romaji:

<|phoneme_start|>EH1 N JH AH0 N IH0 R<|phoneme_end|>

To switch language mid-line, wrap only the words that change language in a language tag, for example:

Thanks for coming. <french>Je voudrais une table pour deux, s'il vous plaît.</french>

Fish Audio labels language tags experimental: they are not guaranteed for every language or voice, and the model may switch language on its own. Voices cloned from a speaker fluent in both languages work best.

Code examples

# ─── 一个音色,带表演指示 ─────────────────────────────────
curl -X POST https://api.apimodels.app/v1/audio/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "fish-drama-3",
    "text": "[barely holding it together, forcing a smile] I am fine. Really. <whisper>Go on without me.</whisper>",
    "reference_id": "VOICE_ID",
    "format": "mp3"
  }'
# { "code": 200, "data": { "taskId": "clxxx", "state": "pending" } }


# ─── 两个人的对白:reference_id 传数组,<|speaker:N|> 切换 ─────
curl -X POST https://api.apimodels.app/v1/audio/generations \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "fish-drama-3",
    "text": "<|speaker:0|>[tired, keeping her voice down] Did you lock the door?<|speaker:1|>[too quickly] Of course I did.<|speaker:0|>[a long pause] Then why is it open?",
    "reference_id": ["VOICE_ID_A", "VOICE_ID_B"],
    "format": "wav"
  }'


# ─── 取结果 ─────────────────────────────────────────────
curl "https://api.apimodels.app/v1/audio/generations?task_id=TASK_ID" \
  -H "Authorization: Bearer YOUR_API_KEY"
# { "code": 200, "data": { "state": "completed", "resultUrls": ["https://r2.apimodels.app/..."] } }

Request parameters

The other S2.1 Pro parameters (prosody, mp3_bitrate, sample_rate, latency, chunk_length) are passed to the model unchanged; ranges are listed in the S2.1 Pro docs.

Drama 3 or S2.1 Pro

• Use Drama 3 when the line has to be performed: emotional dialogue, audio drama, game characters, several speakers in one scene.
• Use S2.1 Pro for plain narration that needs no direction: $0.03 per 1,000 bytes, with an evaluation tier as well.
• In our test, 1,000 English characters took 9.6 seconds and returned 89 seconds of audio; 500 Chinese characters took 13.4 seconds for 146 seconds of audio. These are whole-passage render times, not time to first audio.

Notes

• Async: create a task for a taskId, then poll GET ?task_id= for the result.
• Cloned voices are shared with S2.1 Pro, and your first 20 are free; see the S2.1 Pro docs.
• Result files are deleted after 30 days; download or re-host them in time.
• Charged only on success.

Try it in the Playground