Difference between revisions of "2026:Lyrics Transcription Results"

From MIREX Wiki
(Create results page ALT)
 
(→‎Submissions: PDF links (abstracts archived on futuremirex.com))
 
Line 10: Line 10:
 
|-
 
|-
 
| Frangiadaki et al.
 
| Frangiadaki et al.
−
|  
+
| [http://futuremirex.com/portal/wp-content/uploads/2026/lyrics-transcription/Frangiadaki.pdf PDF]
 
| NA
 
| NA
 
| ~132.15B
 
| ~132.15B
Line 17: Line 17:
 
|-
 
|-
 
| UFAL-PMCG-Whisper
 
| UFAL-PMCG-Whisper
−
|
+
| [http://futuremirex.com/portal/wp-content/uploads/2026/lyrics-transcription/UFAL-PMCG-Whisper.pdf PDF]
 
| NA
 
| NA
 
| 1.55B
 
| 1.55B
Line 24: Line 24:
 
|-
 
|-
 
| UFAL-PMCG-Qwen-ASR
 
| UFAL-PMCG-Qwen-ASR
−
|
+
| [http://futuremirex.com/portal/wp-content/uploads/2026/lyrics-transcription/UFAL-PMCG-Qwen-ASR.pdf PDF]
 
| NA
 
| NA
 
| 3.6B
 
| 3.6B
Line 31: Line 31:
 
|-
 
|-
 
| UFAL-PMCG-Whisper+Qwen
 
| UFAL-PMCG-Whisper+Qwen
−
|
+
| [http://futuremirex.com/portal/wp-content/uploads/2026/lyrics-transcription/UFAL-PMCG-Whisper-Qwen.pdf PDF]
 
| NA
 
| NA
 
| 11.8B
 
| 11.8B
Line 38: Line 38:
 
|-
 
|-
 
| UFAL-PMCG-Gemma
 
| UFAL-PMCG-Gemma
−
|
+
| [http://futuremirex.com/portal/wp-content/uploads/2026/lyrics-transcription/UFAL-PMCG-Gemma.pdf PDF]
 
| NA
 
| NA
 
| 11.6B
 
| 11.6B

Latest revision as of 10:21, 14 October 2026

Submissions

Team PDF Training size Number of parameters Training hours Inference real-time factor
Frangiadaki et al. PDF NA ~132.15B NA 1.20
UFAL-PMCG-Whisper PDF NA 1.55B NA 0.0386
UFAL-PMCG-Qwen-ASR PDF NA 3.6B NA 0.0969
UFAL-PMCG-Whisper+Qwen PDF NA 11.8B NA 0.1753
UFAL-PMCG-Gemma PDF NA 11.6B NA 0.2111

Results

The results below are aggregate scores for each dataset. Lower is better for CER, WER, and WER'; higher is better for the F-scores. The formatting-token scores are F-scores for Punctuation (P), Parentheses (B), Line Breaks (L), and Section Breaks (S). Because all submissions were multilingual natively, we directly report complete results on the data.

Jam-ALT

Team CER WER WER' F-score (P) F-score (B) F-score (L) F-score (S)
Frangiadaki et al.(1) 0.0881 0.1470 0.1913 0.5134 0.4932 0.8026 0.7587
UFAL-PMCG-Whisper 0.1991 0.2860 0.3317 0.4295 NA 0.7442 0.0000
UFAL-PMCG-Qwen-ASR 0.1598 0.2593 0.3190 0.2584 NA 0.5438 NA
UFAL-PMCG-Whisper+Qwen 0.1714 0.2622 0.3162 0.3210 NA 0.6764 0.2661
UFAL-PMCG-Gemma 0.1785 0.2806 0.3301 0.1447 0.0943 0.8118 0.2431

(1) The lyrics for the Spanish songs were pre-computed by the authors, because inference ran OOM on a V100-32GB.

MUSDB-ALT

Team CER WER WER' F-score (P) F-score (B) F-score (L) F-score (S)
Frangiadaki et al.(2) 0.1163 0.1617 0.1970 0.2941 0.3190 0.8378 0.7485
UFAL-PMCG-Whisper 0.2031 0.2740 0.3126 0.2335 NA 0.7748 0.2159
UFAL-PMCG-Qwen-ASR 0.1569 0.2181 0.2697 0.1174 NA 0.5615 NA
UFAL-PMCG-Whisper+Qwen 0.1765 0.2362 0.2858 0.1563 NA 0.6896 0.2430
UFAL-PMCG-Gemma 0.2310 0.3123 0.3520 0.0680 0.1389 0.8138 0.2331

(2) The lyrics were pre-computed by the authors, because inference ran OOM on a V100-32GB.

JamendoLyrics++

JamendoLyrics++ was used as an additional evaluation set. It contains 80 English songs that do not overlap the previous 2 datasets. The lyrics however do not contain punctuation or capitalization, so only a reduced set of metrics is used.

Team CER WER F-score (L) F-score (S)
Frangiadiaki et al. 0.1263 0.1768 0.8250 0.7655
UFAL-PMCG-Whisper 0.1835 0.2625 0.7495 0.2500
UFAL-PMCG-Qwen-ASR 0.1391 0.2144 0.5562 NA
UFAL-PMCG-Whisper+Qwen 0.1747 0.2495 0.6698 0.2765
UFAL-PMCG-Gemma 0.2429 0.3353 0.7667 0.2269