āĻ…āĻĄāĻŋāĻ“ āĻāφāχ āĻ—āĻžāχāĻĄ

āĻ…āĻĄāĻŋāĻ“ āĻ‰ā§ŽāϏ āĻŦāĻŋāĻ­āĻžāϜāύ⧇āϰ āϜāĻ¨ā§āϝ SI-SDR

āĻ¸ā§āϕ⧇āϞ-āχāύāĻ­āĻžāϰāĻŋāϝāĻŧ⧇āĻ¨ā§āϟ āϏāĻŋāĻ—āĻ¨ā§āϝāĻžāϞ-āϟ⧁-āĻĄāĻŋāĻ¸ā§āϟāĻ°ā§āĻļāύ āϰ⧇āĻļāĻŋāĻ“, āĻŦāĻž SI-SDR, āĻāĻ•āϟāĻŋ āĻ…āύ⧁āĻŽāĻŋāϤ āĻ…āĻĄāĻŋāĻ“ āϏ⧋āĻ°ā§āϏāϕ⧇ āĻāĻ•āϟāĻŋ āϰ⧇āĻĢāĻžāϰ⧇āĻ¨ā§āϏ⧇āϰ āϏāĻžāĻĨ⧇ āϤ⧁āϞāύāĻž āĻ•āϰ⧇, āĻāĻ•āϟāĻŋ āϏāĻžāĻŽāĻ—ā§āϰāĻŋāĻ• āϗ⧇āχāύ āϏāĻŽāĻ¨ā§āĻŦāϝāĻŧ āĻ…āύ⧁āĻŽā§‹āĻĻāύ⧇āϰ āĻĒāϰ⧇āĨ¤

āĻāχ āύāĻŋāĻ°ā§āĻĻ⧇āĻļāĻŋāĻ•āĻž āĻļ⧁āύ⧁āύāĻāφāχ-āĻŦāĻ°ā§āĻŖāĻŋāϤ ¡ ~3 min

āĻāχ āĻĒ⧃āĻˇā§āĻ āĻžāϝāĻŧ3 āĻŽāĻŋāύāĻŋāϟ āĻĒāĻĄāĻŧāĻž āĻšāϝāĻŧ⧇āϛ⧇
  1. āĻ“āĻ­āĻžāϰāĻ­āĻŋāω
  2. āĻ—āĻ­ā§€āϰ āĻĄā§āĻŦ
  3. āĻŦāĻžāĻ¸ā§āϤāĻŦ-āĻŦāĻŋāĻļā§āĻŦ āĻŦāĻžāĻ¸ā§āϤāĻŦāĻžāϝāĻŧāύ
  4. The Future of SI-SDR for Audio Source Separation
  5. āĻāϟāĻžāϕ⧇ āĻŦāĻžāĻ¸ā§āϤāĻŦ⧇ āĻĒā§āϰāϝāĻŧā§‹āĻ— āĻ•āϰāĻž āĻšāĻšā§āϛ⧇
  6. āĻ…āĻ¨ā§āĻŦ⧇āώāĻŖ āϚāĻžāϞāĻŋāϝāĻŧ⧇ āϝāĻžāύ
  7. āĻĒā§āϰāĻžāϝāĻŧāĻļāχ āϜāĻŋāĻœā§āĻžāĻžāϏāĻŋāϤ āĻĒā§āϰāĻļā§āύāĻžāĻŦāϞ⧀

āĻ“āĻ­āĻžāϰāĻ­āĻŋāω

It is a common objective score for source-separation experiments. A higher value means less residual error under that definition, but it does not by itself prove that speech is intelligible, music sounds natural or a model will work on a different recording.

āĻ—āĻ­ā§€āϰ āĻĄā§āĻŦ

Audio source separation estimates one component of a mixture, such as a voice in music, and compares it with an available isolated reference. A metric is needed to summarize how close the estimate is. Le Roux and colleagues proposed scale-invariant SDR after explaining problems with common uses of an older BSS_eval SDR definition, especially for single-channel separation. SI-SDR first allows an overall gain adjustment to the reference and then compares energy aligned with that reference to residual error energy. This makes the score insensitive to simple output-volume changes while still penalizing mismatched waveform content. Scale invariance has a purpose and a cost. If one estimate is exactly a quieter version of the reference, SI-SDR can remain high after the allowed gain fit. That is useful when the task is waveform shape independent of volume, but not enough for a product that must preserve loudness. The alternative scale-dependent SDR retains sensitivity to gain. A benchmark should state which metric implementation and alignment rules it uses rather than saying only “SDR.” Reference timing, channel handling and silent segments can complicate computation. An average score may hide source-specific and song-specific failures. Two separators with similar SI-SDR can sound different to listeners because a brief speech consonant, reverb tail or musical transient may matter disproportionately. The metric is waveform-oriented, not a complete measure of perceived quality or downstream ASR accuracy. Use perceptual listening and task metrics alongside it, especially for accessibility or transcription. For fair comparison, use identical held-out mixtures, the same reference stems, sample rate and score implementation. Report median or distribution as well as mean when a few catastrophic cases matter. Do not tune repeatedly on a published test set and then call its score independent. SI-SDR is a valuable numerical tool when its invariance and limitations are explicit.

āĻŦāĻžāĻ¸ā§āϤāĻŦ-āĻŦāĻŋāĻļā§āĻŦ āĻŦāĻžāĻ¸ā§āϤāĻŦāĻžāϝāĻŧāύ

A researcher compares two dialogue separators with SI-SDR on the same held-out mixtures and reference stems.

An evaluator checks waveform alignment before interpreting a very poor score on an otherwise recognizable sound.

A team listens for metallic artifacts even when a new model improves average SI-SDR.

An engineer reports results for vocals and drums separately rather than hiding one weak source in a mean.

The Future of SI-SDR for Audio Source Separation

Objective separation metrics will become more robust and task-specific, but no one number will capture every artifact a listener notices. Systems may combine SI-SDR with perceptual measures and the performance of a downstream captioner or hearing-accessibility tool. Public benchmarks should publish evaluation code and reference handling so results remain comparable. Product teams need to test level preservation separately when scale matters. For songs or films with sparse, unusual sources, distribution plots and listening samples can reveal failures hidden by average improvement. The best evaluation matches the purpose of the separated audio.

āĻāϟāĻžāϕ⧇ āĻŦāĻžāĻ¸ā§āϤāĻŦ⧇ āĻĒā§āϰāϝāĻŧā§‹āĻ— āĻ•āϰāĻž āĻšāĻšā§āϛ⧇

āϕ⧇āύ āĻāϟāĻž āϗ⧁āϰ⧁āĻ¤ā§āĻŦāĻĒā§‚āĻ°ā§āĻŖ

  • āĻ…ā§āϝāĻžāĻ•ā§āϏ⧇āϏ āĻāĻŦāĻ‚ āĻĒ⧌āρāĻ›āĻžāύ⧋. āĻāϟāĻŋ āĻŸā§āϰāĻžāĻ¨ā§āϏāĻ•ā§āϰāĻŋāĻĒāĻļāύ, āĻŦāĻ°ā§āĻŖāύāĻž āĻāĻŦāĻ‚ āĻ­āϝāĻŧ⧇āϏ āχāĻ¨ā§āϟāĻžāϰāĻĢ⧇āϏ⧇āϰ āĻŽāĻžāĻ§ā§āϝāĻŽā§‡ āĻ…ā§āϝāĻžāĻ•ā§āϏ⧇āϏāϝ⧋āĻ—ā§āϝāϤāĻž āωāĻ¨ā§āύāϤ āĻ•āϰ⧇āĨ¤
  • āĻ–āϰāϚ āĻ“ āĻŦāĻžāĻœā§‡āϟ. āĻŽāĻŋāĻĄāĻŋāϝāĻŧāĻž āĻĻāϞāϗ⧁āϞāĻŋ āϛ⧋āϟ āĻŦāĻžāĻœā§‡āĻŸā§‡āϰ āϏāĻžāĻĨ⧇ āĻĒāĻžāϞāĻŋāĻļ āĻ•āϰāĻž āĻ…āĻĄāĻŋāĻ“ āĻĻā§āϰ⧁āϤ āĻĒāĻžāĻ āĻžāϤ⧇ āĻĒāĻžāϰ⧇āĨ¤
  • āĻ—āϤāĻŋ āĻāĻŦāĻ‚ āĻ¸ā§āϕ⧇āϞ. āĻ—ā§āϰāĻžāĻšāĻ•-āĻŽā§āĻ–ā§€ āϏāĻŋāĻ¸ā§āĻŸā§‡āĻŽāϗ⧁āϞāĻŋ āĻŦ⧃āĻšāĻ¤ā§āϤāϰ āĻ¸ā§āϕ⧇āϞ⧇ āĻ•āĻĨā§āϝ āĻŽāĻŋāĻĨāĻ¸ā§āĻ•ā§āϰāĻŋāϝāĻŧāĻž āĻĒā§āϰāĻ•ā§āϰāĻŋāϝāĻŧāĻž āĻ•āϰāϤ⧇ āĻĒāĻžāϰ⧇āĨ¤

āϜāĻ¨ā§āϝ āϏāϤāĻ°ā§āĻ•

  • āϏāĻŽā§āĻŽāϤāĻŋ āĻ…āύ⧁āĻĒāĻ¸ā§āĻĨāĻŋāϤ āĻšāϞ⧇ āĻ­āϝāĻŧ⧇āϏ āĻ…āĻĒāĻŦā§āϝāĻŦāĻšāĻžāϰ āĻāĻŦāĻ‚ āĻ›āĻĻā§āĻŽāĻŦ⧇āĻļ⧇āϰ āĻā§āρāĻ•āĻŋ āĻŦ⧇āĻĄāĻŧ⧇ āϝāĻžāϝāĻŧāĨ¤
  • āωāĻšā§āϚāĻžāϰāĻŖ, āωāĻĒāĻ­āĻžāώāĻž āĻŦāĻž āϕ⧋āϞāĻžāĻšāϞāĻĒā§‚āĻ°ā§āĻŖ āĻĒāϰāĻŋāĻŦ⧇āĻļ⧇ āϝāĻĨāĻžāĻ°ā§āĻĨāϤāĻž āĻšā§āϰāĻžāϏ āĻĒ⧇āϤ⧇ āĻĒāĻžāϰ⧇āĨ¤
  • āϏāĻŋāĻ¨ā§āĻĨ⧇āϟāĻŋāĻ• āĻ…āĻĄāĻŋāĻ“ āĻĒāϰāĻŋāĻˇā§āĻ•āĻžāϰ āϞ⧇āĻŦ⧇āϞ āĻ›āĻžāĻĄāĻŧāĻž āĻ–āĻžāρāϟāĻŋ āĻŦāĻ•ā§āϤ⧃āϤāĻž āĻšāĻŋāϏāĻžāĻŦ⧇ āϭ⧁āϞ āĻšāϤ⧇ āĻĒāĻžāϰ⧇āĨ¤

āĻĒāϰāĻŦāĻ°ā§āϤ⧀ āĻĒāĻĻāĻ•ā§āώ⧇āĻĒ

  1. āĻ­āϝāĻŧ⧇āϏ āĻ•ā§āϝāĻžāĻĒāϚāĻžāϰ, āĻ•ā§āϞ⧋āύāĻŋāĻ‚ āĻāĻŦāĻ‚ āĻĒ⧁āύāσāĻŦā§āϝāĻŦāĻšāĻžāϰ⧇āϰ āϜāĻ¨ā§āϝ āϏ⧁āĻ¸ā§āĻĒāĻˇā§āϟ āϏāĻŽā§āĻŽāϤāĻŋ āĻĒāĻžāύāĨ¤
  2. āĻŦāĻŋāĻ­āĻŋāĻ¨ā§āύ āĻ¸ā§āĻĒāĻŋāĻ•āĻžāϰ āĻāĻŦāĻ‚ āĻŦā§āϝāĻžāĻ•āĻ—ā§āϰāĻžāωāĻ¨ā§āĻĄ āĻ•āĻ¨ā§āĻĄāĻŋāĻļāύ āϜ⧁āĻĄāĻŧ⧇ āĻŽāĻžāύ āĻĒāϰ⧀āĻ•ā§āώāĻž āĻ•āϰ⧁āύāĨ¤
  3. āĻ•āĻ–āύ āĻāĻ•āϜāύ āĻŽāĻžāύ⧁āώāϕ⧇ āφāωāϟāĻĒ⧁āϟ āĻĒāĻ°ā§āϝāĻžāϞ⧋āϚāύāĻž āĻŦāĻž āĻ…āύ⧁āĻŽā§‹āĻĻāύ āĻ•āϰāϤ⧇ āĻšāĻŦ⧇ āϤāĻž āύāĻŋāĻ°ā§āϧāĻžāϰāĻŖ āĻ•āϰ⧁āύāĨ¤
  4. āϏāĻŋāĻ¨ā§āĻĨ⧇āϟāĻŋāĻ• āĻ…āĻĄāĻŋāĻ“ āϞ⧇āĻŦ⧇āϞ āĻ•āϰ⧁āύ āĻāĻŦāĻ‚ āĻĻāĻžāϝāĻŧāĻŦāĻĻā§āϧāϤāĻžāϰ āϜāĻ¨ā§āϝ āĻŽā§‚āϞ āϰ⧇āĻ•āĻ°ā§āĻĄ āϰāĻžāϖ⧁āύāĨ¤

āĻ…āĻ¨ā§āĻŦ⧇āώāĻŖ āϚāĻžāϞāĻŋāϝāĻŧ⧇ āϝāĻžāύ

āĻĢā§āϰāĻŋ āύāĻŋāωāϜāϞ⧇āϟāĻžāϰ

āĻĒā§āϰāϤāĻŋāĻĻāĻŋāύ⧇āϰ āĻāφāχ āĻŦā§āϰāĻŋāĻĢāĻŋāĻ‚ āĻĒāĻžāύ

āĻĒā§āϰāϤāĻŋ āϏāĻĒā§āϤāĻžāĻšā§‡āϰ āĻ•āĻžāĻœā§‡āϰ āĻĻāĻŋāύ āϏāĻ•āĻžāϞ⧇ āϤāĻŋāύāϟāĻŋ āϝāĻžāϚāĻžāχāĻ•ā§ƒāϤ āĻāφāχ āĻ¸ā§āĻŸā§‹āϰāĻŋ, āϏāĻšāϜ āχāĻ‚āϰ⧇āϜāĻŋāϤ⧇ āϞ⧇āĻ–āĻžāĨ¤ āϚāĻŋāϰāĻ•āĻžāϞ⧇āϰ āϜāĻ¨ā§āϝ āĻĢā§āϰāĻŋ, āϕ⧋āύ⧋ āĻŦāĻŋāĻœā§āĻžāĻžāĻĒāύ āύ⧇āχāĨ¤

āĻĒā§āϰāϤāĻŋ āϏāĻĒā§āϤāĻžāĻšā§‡āϰ āĻ•āĻžāĻœā§‡āϰ āĻĻāĻŋāύ⧇ āĻāĻ•āϟāĻŋ āχāĻŽā§‡āϞāĨ¤ āĻāĻ• āĻ•ā§āϞāĻŋāϕ⧇āχ āφāύāϏāĻžāĻŦāĻ¸ā§āĻ•ā§āϰāĻžāχāĻŦ āĻ•āϰ⧁āύāĨ¤ āφāĻŽāϰāĻž āĻ•āĻ–āύ⧋āχ āφāĻĒāύāĻžāϰ āĻ āĻŋāĻ•āĻžāύāĻž āĻŦāĻŋāĻ•ā§āϰāĻŋ āĻŦāĻž āĻļā§‡ā§ŸāĻžāϰ āĻ•āϰāĻŋ āύāĻžāĨ¤

āύāĻŋāĻœā§‡āϕ⧇ āĻĒāϰ⧀āĻ•ā§āώāĻž āĻ•āϰ⧁āύ

Take the SI-SDR for Audio Source Separation quiz

āĻĒā§āϰāϤāĻŋāϟāĻŋ āωāĻ¤ā§āϤāϰ⧇āϰ āϤāĻžā§ŽāĻ•ā§āώāĻŖāĻŋāĻ• āĻĢāĻŋāĻĄāĻŦā§āϝāĻžāĻ•, āĻāĻŦāĻ‚ āϕ⧋āĻ°ā§āϏ āĻĒāĻžāϏ āĻ•āϰāĻžāϰ āĻĒāϰ āϝāĻžāϚāĻžāχāϝ⧋āĻ—ā§āϝ āφāχāĻĄāĻŋ āϏāĻš āĻāĻ•āϟāĻŋ āĻļā§‡ā§ŸāĻžāϰāϝ⧋āĻ—ā§āϝ āϏāĻžāĻ°ā§āϟāĻŋāĻĢāĻŋāϕ⧇āϟāĨ¤

āϕ⧁āχāϜ āĻļ⧁āϰ⧁ āĻ•āϰ⧁āύ

āĻĢā§āϰāĻŋ āĻāφāχ āĻļāĻŋāĻ•ā§āώāĻžāϕ⧇ āϏāĻŽāĻ°ā§āĻĨāύ āĻ•āϰ⧁āύāĨ¤ AI Understanding āĻāĻ•āϟāĻŋ 501(c)(3) āĻ…āϞāĻžāĻ­āϜāύāĻ• āĻĒā§āϰāϤāĻŋāĻˇā§āĻ āĻžāύ — āϕ⧋āύ⧋ āĻŦāĻŋāĻœā§āĻžāĻžāĻĒāύ āύ⧇āχ, āϕ⧋āύ⧋ āĻĒ⧇āĻ“ā§ŸāĻžāϞ āύ⧇āχ, āĻ•āĻ–āύ⧋āχ āύāĻžāĨ¤ āĻāĻ•āϟāĻŋ āĻĻāĻžāύ āĻ•āϰ⧁āύ

āĻĒā§āϰāĻžāϝāĻŧāĻļāχ āϜāĻŋāĻœā§āĻžāĻžāϏāĻŋāϤ āĻĒā§āϰāĻļā§āύāĻžāĻŦāϞ⧀

What is SI-SDR for Audio Source Separation?

Scale-invariant signal-to-distortion ratio, or SI-SDR, compares an estimated audio source with a reference after allowing one overall gain adjustment. It is a common objective score for source-separation experiments. A higher value means less residual error under that definition, but it does not by itself prove that speech is intelligible, music sounds natural or a model will work on a different recording.

What are real examples of SI-SDR for Audio Source Separation in practice?

A researcher compares two dialogue separators with SI-SDR on the same held-out mixtures and reference stems. An evaluator checks waveform alignment before interpreting a very poor score on an otherwise recognizable sound. A team listens for metallic artifacts even when a new model improves average SI-SDR. An engineer reports results for vocals and drums separately rather than hiding one weak source in a mean.

What is next for SI-SDR for Audio Source Separation?

Objective separation metrics will become more robust and task-specific, but no one number will capture every artifact a listener notices. Systems may combine SI-SDR with perceptual measures and the performance of a downstream captioner or hearing-accessibility tool. Public benchmarks should publish evaluation code and reference handling so results remain comparable. Product teams need to test level preservation separately when scale matters. For songs or films with sparse, unusual sources, distribution plots and listening samples can reveal failures hidden by average improvement. The best evaluation matches the purpose of the separated audio.

A separator improves average SI-SDR but loses consonants. What remains to evaluate?

Waveform ratios do not capture every important perceptual effect.

āĻļāĻŋāĻ–āϤ⧇ āĻĨāĻžāϕ⧁āύ

āϏāĻŽā§āĻĒāĻ°ā§āĻ•āĻŋāϤ āĻ—āĻžāχāĻĄ

āĻāχ āĻŦāĻŋāώāϝāĻŧ⧇āϰ āϜāĻ¨ā§āϝ āĻŦāĻžāĻ›āĻžāχ āĻ•āϰāĻž āφāϰāĻ“ āĻ—āĻžāχāĻĄ

āĻĒāϰāĻŦāĻ°ā§āϤ⧀ āφāĻĒāĻĒāϰāĻŦāĻ°ā§āϤ⧀ āĻ—āĻžāχāĻĄ

Demucs āϏāĻ™ā§āĻ—ā§€āϤ āĻ‰ā§ŽāϏ āĻŦāĻŋāĻšā§āϛ⧇āĻĻ

āĻ…āĻĄāĻŋāĻ“ āĻāφāχ

āĻāĻ•āϟāĻŋ āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ AI āĻļāĻ‚āϏāĻžāĻĒāĻ¤ā§āϰ āĻĒāĻžāύāĻāĻ•āϟāĻŋ āϏāĻ‚āĻ•ā§āώāĻŋāĻĒā§āϤ āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ āϕ⧋āĻ°ā§āϏ āĻāĻŦāĻ‚ āĻĒāϰ⧀āĻ•ā§āώāĻž. āϕ⧋āύ⧋ āĻ–āϰāϚ āύ⧇āχ, āϕ⧋āύ⧋ āĻ•āĻžāĻ°ā§āĻĄ āύ⧇āχāĨ¤ āĻĒā§āϰāĻžāϝāĻŧ 2.5 āϘāĻ¨ā§āϟāĻžāĨ¤āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ āĻļ⧁āϰ⧁ āĻ•āϰ⧁āύ →