Glossary
Beam search is the standard decoding algorithm used by sequence models like Whisper. Instead of greedily picking the single most-likely next word at each step, the algorithm tracks the top k (the "beam width") candidate sequences and picks the best-scoring complete path. It's the reason ASR doesn't make obvious greedy mistakes like "I scream" instead of "ice cream" — by considering multiple paths, the right one wins on cumulative score.
Greedy vs beam search
Greedy decoding picks the highest-probability next token at every step. Fast but myopic — early bad choices propagate. Beam search with beam width k=5 tracks the top 5 partial sequences after each step. After processing the full audio, it returns the highest-scoring complete sequence. Roughly 5x more compute than greedy, in exchange for noticeably better accuracy on ambiguous audio.
Beam width tradeoff
Higher beam width = better quality up to a point, then diminishing returns. k=1 is greedy. k=5 is the practical sweet spot for ASR — most accuracy gain is captured here. k=10-20 recovers a few more cases but costs proportionally more. Beyond k=20, you're spending compute on long-tail cases that humans get wrong too.
Length penalty and other adjustments
Pure beam search prefers shorter sequences (each token has probability < 1, so short sequences accumulate higher product probability). Length-normalization or a length penalty term corrects this. Whisper uses temperature sampling at decoding time — when token probabilities are too uniform (model is uncertain), it lowers temperature and re-decodes. We use the standard Whisper decoding strategy without modification — proprietary tweaks tend to overfit.
In practice
Audio: "I scream for ice cream." After processing the first three words, greedy decoding might pick "I scream for I scream" (misinterpreting the last "ice cream" as another "I scream" because they sound similar). Beam search keeps both "I scream" and "ice cream" as candidates after each chunk, and the language model's preference for "ice cream" in the food context wins on cumulative score.
Related terms
Further reading
See accurate decoding in practice
Get started — freeLast updated: September 20, 2026
Frequently Asked Questions
What does Beam search mean in transcription?
A decoding strategy used by ASR engines: instead of picking the single most-likely next word at each step (greedy), the system tracks the top-k candidate sequences and picks the highest-scoring complete path. Trades compute for accuracy.
Why does Beam search matter when choosing a transcription workflow?
Beam search affects how you understand transcript quality, timing, compatibility, or the technology behind speech-to-text results.
Where can I apply Beam search on Transcript.you?
Use the related workflow at /ai-transcription when you want to see how this glossary concept connects to an actual transcription task.
Browse the full transcription glossary or read the complete guide to AI transcription.