Fish Audio: S1

S1 is Fish Audio’s previous-generation text-to-speech model for expressive narration and single-speaker voice output. It reads a supplied script and returns audio, using a selected or reference-conditioned voice. The full S1 model is distinct from S1-mini, the smaller distilled model; the producer introduces them as 4-billion- and 0.5-billion-parameter variants respectively. This profile covers the named S1 model.

For narration, prepare the words that should actually be spoken and mark the places where the delivery should change. A useful explainer script states its main point first, develops one idea at a time and uses punctuation to separate phrases that need a natural pause. For a short instruction, keep the wording clear when heard without a screen. Preserve the intended reading of names, abbreviations and units, and check numbers in the spoken output rather than assuming a visually readable string has only one pronunciation.

S1 uses emotion and delivery cues in parentheses, such as (angry), (sad), (in a hurry) or (chuckling). Place a cue before the sentence or phrase it should influence, with actual speech after it. The delivery belongs to the text that follows, so a mood change should appear where the performance needs to change. This is a different syntax from S2’s square-bracket directions; keep the script written for the selected generation instead of assuming the two markup forms are interchangeable.

A voice can be selected through a saved voice identifier or conditioned from a reference recording together with its transcript. For an instant reference workflow, the general speech guide recommends a clean 10–30-second sample. Choose a sample whose speaking style and recording quality fit the intended result, and make the transcript match the words in that sample. For a recurring narrator, reuse the same voice reference across passages so a change of subject does not unintentionally become a change of character.

The S1 release names English, Chinese, Japanese, German, French, Spanish, Korean, Arabic, Russian, Dutch, Italian, Polish and Portuguese among its supported languages and describes mixing languages within a script. Supply the target-language wording rather than treating speech synthesis as a translation step. For a passage with a borrowed name or a language switch, keep the spelling and surrounding sentence context clear and review how the selected voice handles that transition.

S1’s producer interface supports one speaker per synthesis request. A dialogue can be prepared as separate segments with their own voice references, then assembled by the application; S2-family native multi-speaker synthesis is a separate capability. Keep the role, voice and text associated with each segment when preparing character lines. The synthesis result is an audio performance of the script, while timing an exchange or combining it with music and video is a later editing step.

Output options include MP3, WAV, raw PCM and Opus, with MP3 as the default. The documented output is mono; WAV and PCM use 16-bit samples, and format-specific sample-rate choices differ. MP3 supports 64, 128 and 192 kbps, while Opus uses 48 kHz. Speed can be adjusted from 0.5 to 2.0 with volume controls. Choose an encoded file for a voiceover artifact or an appropriate stream for playback, and retain the selected voice and script with the result.

The producer processes long text in chunks, with a documented chunk-length range of 100–300 and an audio-token budget per chunk rather than a guaranteed overall recording duration. Text normalisation can expand English and Chinese numbers into natural reading, and S1’s repetition penalty can reduce repeated audio patterns. Output loudness normalisation is a different setting and has no effect on S1. Inspect pronunciation, phrase boundaries, repeated sounds and the intended emotional shift before assembling the finished passage, then revise the specific line or reference that needs attention.