topabaem commited on
Commit
53536c4
Β·
verified Β·
1 Parent(s): 116f2a9

Card: speed on the demo Space (source 86b6e7cbb263)

Browse files
Files changed (1) hide show
  1. README.md +3 -1
README.md CHANGED
@@ -76,6 +76,8 @@ print(stt.transcribe("μŒμ„±.wav")) # 파일 경둜, λ˜λŠ” stt.transcribe(λ°°
76
 
77
  Apple M5 Proμ—μ„œ 60λΆ„ μŒμ„±μ„ Core AI둜 5.50μ΄ˆμ— λ°›μ•„μΌμŠ΅λ‹ˆλ‹€(μ‹€μ‹œκ°„μ˜ 654λ°°, μ΅œλŒ€ λ©”λͺ¨λ¦¬ 3.4GB). 같은 λ§₯μ—μ„œ PyTorch MPSλŠ” 21.7μ΄ˆμ˜€μŠ΅λ‹ˆλ‹€. 이 값은 ꡬ쑰와 좜λ ₯ 크기가 같은 이전 κ°€μ€‘μΉ˜λ‘œ 잰 κ°’μž…λ‹ˆλ‹€.
78
 
 
 
79
  ## ν•œκ³„
80
 
81
  * **λ‚­λ…μ²΄λ‘œλ§Œ ν‰κ°€ν–ˆμŠ΅λ‹ˆλ‹€.** λŒ€ν™”, 작음, μ—¬λŸ¬ μ‚¬λžŒμ΄ ν•¨κ»˜ λ§ν•˜λŠ” μŒμ„±μ˜ ν’ˆμ§ˆμ€ μž¬μ§€ μ•Šμ•˜μŠ΅λ‹ˆλ‹€.
@@ -113,7 +115,7 @@ https://huggingface.co/Haverbex/eg2-ko-reader/discussions/new
113
  * **What it is**: a Korean syllable CTC reader (6-layer Transformer, 19.7M parameters) that reads the output of the frozen EmbeddingGemma 2 audio encoder (25 tokens per second, 512 numbers each). No autoregressive decoder and no language model.
114
  * **Use**: `hf download Haverbex/eg2-ko-reader --local-dir eg2-ko-reader`, `pip install -r eg2-ko-reader/requirements.txt`, `python eg2-ko-reader/eg2_ko_reader.py audio.wav`. The download is about 1.6 GB and includes the encoder.
115
  * **Accuracy** (character error rate without spaces, greedy): Zeroth-Korean test 7.26 %, FLEURS ko test 5.59 %. Read speech only; one training seed; no confidence interval for the test sets.
116
- * **Speed**: 60 minutes of audio in 5.50 s with Core AI on an Apple M5 Pro (measured with earlier weights of the same architecture).
117
  * **Limits**: Hangul syllables and spaces only (no punctuation, digits or Latin letters); not evaluated on conversation or noise; no streaming; no speaker labels.
118
  * **Training data**: Zeroth-Korean and FLEURS ko (CC BY 4.0), and about 460 hours of YODAS2 Korean (YouTube CC BY video, CC BY 3.0) with machine-made transcripts. Not legally reviewed.
119
  * **License**: the reader, the code and the Core AI bundle are under the EG2-KO Reader License 1.0: individuals may use them for any purpose; organizations may use them for non-commercial purposes and must ask for written permission before commercial use. This is source-available, not an OSI open-source license. `encoder/` is Apache-2.0.
 
76
 
77
  Apple M5 Proμ—μ„œ 60λΆ„ μŒμ„±μ„ Core AI둜 5.50μ΄ˆμ— λ°›μ•„μΌμŠ΅λ‹ˆλ‹€(μ‹€μ‹œκ°„μ˜ 654λ°°, μ΅œλŒ€ λ©”λͺ¨λ¦¬ 3.4GB). 같은 λ§₯μ—μ„œ PyTorch MPSλŠ” 21.7μ΄ˆμ˜€μŠ΅λ‹ˆλ‹€. 이 값은 ꡬ쑰와 좜λ ₯ 크기가 같은 이전 κ°€μ€‘μΉ˜λ‘œ 잰 κ°’μž…λ‹ˆλ‹€.
78
 
79
+ 데λͺ¨ Space(Hugging Face 무료 CPU, `cpu-basic`)μ—μ„œλŠ” 112.1초 μŒμ„±μ„ 21.2μ΄ˆμ—, 4.7초 μŒμ„±μ„ 0.8μ΄ˆμ— λ°›μ•„μΌμŠ΅λ‹ˆλ‹€(μ‹€μ‹œκ°„μ˜ μ•½ 5λ°°, 2026-11-17, ν•œ λ²ˆμ”© 잰 κ°’).
80
+
81
  ## ν•œκ³„
82
 
83
  * **λ‚­λ…μ²΄λ‘œλ§Œ ν‰κ°€ν–ˆμŠ΅λ‹ˆλ‹€.** λŒ€ν™”, 작음, μ—¬λŸ¬ μ‚¬λžŒμ΄ ν•¨κ»˜ λ§ν•˜λŠ” μŒμ„±μ˜ ν’ˆμ§ˆμ€ μž¬μ§€ μ•Šμ•˜μŠ΅λ‹ˆλ‹€.
 
115
  * **What it is**: a Korean syllable CTC reader (6-layer Transformer, 19.7M parameters) that reads the output of the frozen EmbeddingGemma 2 audio encoder (25 tokens per second, 512 numbers each). No autoregressive decoder and no language model.
116
  * **Use**: `hf download Haverbex/eg2-ko-reader --local-dir eg2-ko-reader`, `pip install -r eg2-ko-reader/requirements.txt`, `python eg2-ko-reader/eg2_ko_reader.py audio.wav`. The download is about 1.6 GB and includes the encoder.
117
  * **Accuracy** (character error rate without spaces, greedy): Zeroth-Korean test 7.26 %, FLEURS ko test 5.59 %. Read speech only; one training seed; no confidence interval for the test sets.
118
+ * **Speed**: 60 minutes of audio in 5.50 s with Core AI on an Apple M5 Pro (measured with earlier weights of the same architecture). On the free-CPU demo Space, 112.1 s of audio took 21.2 s (about 5x realtime, measured once, 2026-11-17).
119
  * **Limits**: Hangul syllables and spaces only (no punctuation, digits or Latin letters); not evaluated on conversation or noise; no streaming; no speaker labels.
120
  * **Training data**: Zeroth-Korean and FLEURS ko (CC BY 4.0), and about 460 hours of YODAS2 Korean (YouTube CC BY video, CC BY 3.0) with machine-made transcripts. Not legally reviewed.
121
  * **License**: the reader, the code and the Core AI bundle are under the EG2-KO Reader License 1.0: individuals may use them for any purpose; organizations may use them for non-commercial purposes and must ask for written permission before commercial use. This is source-available, not an OSI open-source license. `encoder/` is Apache-2.0.