emilstabil commited on
Commit
e85f4a9
·
verified ·
1 Parent(s): d9e7f6e

Describe clear-speech and dictation use

Browse files
Files changed (1) hide show
  1. README.md +7 -1
README.md CHANGED
@@ -98,6 +98,11 @@ word timestamps. It is a FastConformer RNN-T fine-tuned from
98
  `nvidia/parakeet-rnnt-110m-da-dk` and runs locally. Its raw output is lowercase
99
  text; pair it with `RyeAI/ekko-pnc` for capitalization and punctuation.
100
 
 
 
 
 
 
101
  ## Benchmark
102
 
103
  **12.37% mean WER** across 26,780 rows from five Danish test sets, evaluated
@@ -214,7 +219,8 @@ pseudo-labelled audio.
214
  - The 44-token vocabulary does not contain uppercase or punctuation tokens.
215
  - Punctuation and capitalization require a separate restoration model such as
216
  `RyeAI/ekko-pnc`.
217
- - Spontaneous conversation remains the hardest benchmark domain.
 
218
 
219
  ## License
220
 
 
98
  `nvidia/parakeet-rnnt-110m-da-dk` and runs locally. Its raw output is lowercase
99
  text; pair it with `RyeAI/ekko-pnc` for capitalization and punctuation.
100
 
101
+ **Best fit:** clear Danish speech from one speaker, delivered at a steady pace,
102
+ such as dictation or reading aloud. The published results are stronger on read
103
+ and formal speech than on spontaneous conversation. Dictation itself was not
104
+ evaluated as a separate test set.
105
+
106
  ## Benchmark
107
 
108
  **12.37% mean WER** across 26,780 rows from five Danish test sets, evaluated
 
219
  - The 44-token vocabulary does not contain uppercase or punctuation tokens.
220
  - Punctuation and capitalization require a separate restoration model such as
221
  `RyeAI/ekko-pnc`.
222
+ - Spontaneous conversation remains the hardest benchmark domain (24.15% WER).
223
+ - Noisy audio and overlapping speakers were not evaluated separately.
224
 
225
  ## License
226