Available now

Stable Audio 3 Medium Base: text to audio

Stable Audio 3 Medium Base makes music and sound from a text description, up to 380 seconds long, and can rework, repair or extend a recording you upload.

Model ID

Input

Fields marked * are required. Review the price before generating.

Duration of the generated audio in seconds. The medium model supports up to 380 seconds (~6m20s).

Text description of the audio to generate.

Example prompt. Edit it or write your own.

Additional settings

Text description of qualities to avoid in the output.

Number of sampling steps. Base (non-distilled) checkpoints typically need ~50 for good quality.

Classifier-free guidance scale. Higher values follow the prompt more strictly; ~7.0 is a good starting point for base checkpoints.

Random seed for reproducible outputs. Omit for a random seed.

If True, the prompt will be expanded using an LLM for more detailed and higher quality results.

Container format for the generated audio output.

Audio bitrate for compressed output formats (e.g., mp3, aac, opus). Format e.g. '192k' or '320k'. Ignored for lossless formats (wav, flac).

Price$0.08 / run

You sign up first. Your settings are kept.

About this model

Overview

Four modes share the same controls. Text to audio starts from a description. Audio to audio reworks an uploaded file, and a noise level sets how far it may drift from the source. Inpainting regenerates only the stretch between a start and an end time, and outpainting adds new audio before or after the recording.

This is the unmodified base version, meant as a starting point for fine-tuning. It runs 50 steps by default and follows the guidance setting. A negative prompt lists what to avoid, and output comes as MP3, WAV, FLAC, OGG, Opus, M4A or AAC.

What you can build

  • Long background tracks for videos and streams
  • Ambient beds and soundscapes
  • A fixed passage in an otherwise good take
  • Intros and endings added to an existing track

Prompting tips

  • Name the genre, instruments, tempo and mood: "ambient electronica, 90 BPM, soft pads, calm".
  • In audio to audio, a noise level near 0.1 stays close to the source; near 1.0 it becomes a new piece.
  • For inpainting, set the start and end a little wider than the problem so the new part blends in.
  • Keep about 50 steps on this version; far fewer give rough results.

What to check

Listen to the complete output, including edit boundaries, stereo balance, noise and the ending.

Inputs

  • Negative Prompt (optional): Text description of qualities to avoid in the output.
  • Duration (optional): Duration of the generated audio in seconds. The medium model supports up to 380 seconds (~6m20s).
  • Number of Inference Steps (optional): Number of sampling steps. Base (non-distilled) checkpoints typically need ~50 for good quality. From 1 to 100.
  • Guidance Scale (optional): Classifier-free guidance scale. Higher values follow the prompt more strictly; ~7.0 is a good starting point for base checkpoints. From 0 to 25.
  • Seed (optional): Random seed for reproducible outputs. Omit for a random seed.
  • Enable Prompt Expansion (optional): If True, the prompt will be expanded using an LLM for more detailed and higher quality results.
  • Output Format (optional): Container format for the generated audio output. Options: mp3, wav, flac, ogg, opus, m4a, aac.
  • Bitrate (optional): Audio bitrate for compressed output formats (e.g., mp3, aac, opus). Format e.g. '192k' or '320k'. Ignored for lossless formats (wav, flac).
  • Prompt (required): Text description of the audio to generate.

Output and limits

  • Length: 1 second to 380 seconds.
  • File formats: MP3, WAV, FLAC, OGG, Opus, M4A, AAC.
  • Sung or spoken words can come out unclear or differ from your text.
  • Exact length, tempo and key are not guaranteed; listen to the whole result before you use it.
  • Prompts that name real artists or protected songs can be refused.
  • Some prompts or outputs can be refused by safety checks.

Price guide

  • up to, default options: $0.08 / run