Commit
·
214cbd9
1
Parent(s):
6bcad9e
Update README.md
Browse files
README.md
CHANGED
|
@@ -26,19 +26,48 @@ model = Speech2Text.from_pretrained(
|
|
| 26 |
"TalTechNLP/espnet2_estonian"
|
| 27 |
)
|
| 28 |
|
|
|
|
|
|
|
| 29 |
speech, rate = soundfile.read("speech.wav")
|
|
|
|
| 30 |
text, *_ = model(speech)
|
|
|
|
| 31 |
```
|
| 32 |
|
| 33 |
#### Limitations and bias
|
| 34 |
|
| 35 |
## Training data
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
|
| 37 |
## Training procedure
|
| 38 |
|
|
|
|
| 39 |
|
| 40 |
## Evaluation results
|
| 41 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 42 |
|
| 43 |
|
| 44 |
### BibTeX entry and citation info
|
|
|
|
| 26 |
"TalTechNLP/espnet2_estonian"
|
| 27 |
)
|
| 28 |
|
| 29 |
+
# read a sound file with 16k sample rate
|
| 30 |
+
import soundfile
|
| 31 |
speech, rate = soundfile.read("speech.wav")
|
| 32 |
+
assert rate == 16000
|
| 33 |
text, *_ = model(speech)
|
| 34 |
+
print(text[0])
|
| 35 |
```
|
| 36 |
|
| 37 |
#### Limitations and bias
|
| 38 |
|
| 39 |
## Training data
|
| 40 |
+
Acoustic training data:
|
| 41 |
+
|
| 42 |
+
| Type | Amount |
|
| 43 |
+
|-----------------------|:------:|
|
| 44 |
+
| Broadcast speech | 591 |
|
| 45 |
+
| Spontaneous speech | 53 |
|
| 46 |
+
| Elderly speech corpus | 53 |
|
| 47 |
+
| Talks, lectures | 49 |
|
| 48 |
+
| Parliament speeches | 31 |
|
| 49 |
+
| *Total* | *761* |
|
| 50 |
+
|
| 51 |
+
Language model training data:
|
| 52 |
+
* Estonian National Corpus 2019
|
| 53 |
+
* OpenSubtitles
|
| 54 |
+
* Speech transcripts
|
| 55 |
|
| 56 |
## Training procedure
|
| 57 |
|
| 58 |
+
Standard EspNet2 Conformer recipe.
|
| 59 |
|
| 60 |
## Evaluation results
|
| 61 |
|
| 62 |
+
### WER
|
| 63 |
+
|
| 64 |
+
|dataset|Snt|Wrd|Corr|Sub|Del|Ins|Err|S.Err|
|
| 65 |
+
|---|---|---|---|---|---|---|---|---|
|
| 66 |
+
|decode_asr_lm_lm_large_valid.loss.ave_5best_asr_model_valid.acc.ave/aktuaalne2021.testset|2864|56575|93.1|4.5|2.4|2.0|8.9|63.4|
|
| 67 |
+
|decode_asr_lm_lm_large_valid.loss.ave_5best_asr_model_valid.acc.ave/jutusaated.devset|273|4677|93.9|3.6|2.4|1.2|7.3|46.5|
|
| 68 |
+
|decode_asr_lm_lm_large_valid.loss.ave_5best_asr_model_valid.acc.ave/jutusaated.testset|818|11093|94.7|2.7|2.5|0.9|6.2|45.0|
|
| 69 |
+
|decode_asr_lm_lm_large_valid.loss.ave_5best_asr_model_valid.acc.ave/www-trans.devset|1207|13865|82.3|8.5|9.3|3.4|21.2|74.1|
|
| 70 |
+
|decode_asr_lm_lm_large_valid.loss.ave_5best_asr_model_valid.acc.ave/www-trans.testset|1648|22707|86.4|7.6|6.0|2.5|16.1|75.7|
|
| 71 |
|
| 72 |
|
| 73 |
### BibTeX entry and citation info
|