Fish Audio: S2.1 Pro Free (free)

S2.1 Pro Free is Fish Audio’s developer-oriented edition of the S2.1 text-to-speech model. It shares the underlying model and core synthesis capabilities with S2.1 Pro: expressive multilingual speech, natural-language performance directions and native multi-speaker dialogue. Its role is testing, prototyping, development and smaller application workflows.

For a prototype voice assistant, begin with a complete sample reply and evaluate how it sounds before testing progressively arriving text. For a learning product, prepare a short explanation with the exact terms and number readings that the listener needs. For a dialogue prototype, define the speakers and their voice references before adding performance directions. These steps develop an audio artifact and an application flow around it rather than making a synthesis request responsible for the whole conversation.

Write the words that should actually be spoken, then place an instruction immediately before the passage whose delivery should change. Square-bracket directions can express a whisper, a pause, a reaction, emphasis or an open-ended speaking style. Put a mid-sentence change where the performance needs it, and follow descriptive directions with actual speech. A useful explainer brief separates its factual wording from the warmth, pace or emphasis desired for the listener.

For one speaker, select a saved voice identifier or provide a reference recording with its transcript. The producer speech guide recommends a clean 10–30-second sample for instant voice conditioning. Choose a reference with the tone and recording quality appropriate to the script, and preserve the same reference across related passages. The voice reference supplies the speaker’s sound; the text and inline directions supply the content and its performance.

Native multi-speaker synthesis uses markers such as <|speaker:0|> and <|speaker:1|> with an ordered array of voice identifiers or separate reference groups. Keep the numeric speaker labels aligned with that ordering. For an instructional exchange, write each speaker’s line and place the delivery direction after the speaker marker where it belongs. This generates a dialogue within one request; it is a different workflow from S1’s separate single-speaker segments and later assembly.

The producer workflow offers MP3, WAV, raw PCM and Opus output, with MP3 as the default. MP3 supports 64, 128 and 192 kbps; WAV and PCM use documented 16-bit mono sample formats, while Opus uses 48 kHz. Choose the format for a retained narration file, streamed playback or a downstream audio pipeline. Keep the complete script and voice selection with the audio so a later edit can address one line without losing the intended performance.

Speed can be adjusted from 0.5 to 2.0 with volume controls. Text normalisation expands English and Chinese numbers for natural reading. Output loudness normalisation defaults to true and is applied to the S2 family; the same field had no effect on S1. S1’s repetition_penalty does not change S2.1 Pro audio, so an old setting should not be treated as an effective control for this model. Review repeated sounds and wording directly rather than relying on a carried-over parameter to establish the result.

Longer scripts are processed in chunks. The documented chunk-length range is 100–300, and max_new_tokens is an audio-token budget per chunk rather than a total-duration guarantee. Previous-chunk conditioning can carry voice context forward. The producer delivery workflow describes complete audio, incremental audio chunks and live text streaming when the script arrives progressively. Choose the consumption method for the application and preserve enough script context for the next passage’s pronunciation and tone.

For a useful final artifact, listen for omitted or repeated words, number pronunciation, speaker changes, pauses and the intensity of the directions. Different voice references can interpret a cue with different strength, so start with a simple directed line before adding competing tags. In multilingual work, supply the target-language script rather than asking the speech model to invent or translate its content. A focused revision changes the line, cue or reference responsible for the mismatch while retaining the parts of the performance that already fit the brief.