Update README.md
Browse files
README.md
CHANGED
|
@@ -19,31 +19,9 @@ tags:
|
|
| 19 |
|
| 20 |
Canopy-M was trained on Urdu and Sindhi. The numbers below are the evaluation.
|
| 21 |
|
| 22 |
-
| | Urdu `urd` | Sindhi `snd` | Macro |
|
| 23 |
-
|---|---:|---:|---:|
|
| 24 |
-
| CER | 0.132 [0.065, 0.211] | 0.107 [0.091, 0.123] | 0.119 |
|
| 25 |
-
| WER | 0.267 [0.166, 0.383] | 0.271 [0.240, 0.304] | 0.269 |
|
| 26 |
-
| Utterances | 14,681 | 496 | |
|
| 27 |
-
| Speakers | 63 | 164 | |
|
| 28 |
-
|
| 29 |
|
| 30 |
|
| 31 |
|
| 32 |
-
## Model
|
| 33 |
-
|
| 34 |
-
| | |
|
| 35 |
-
|---|---|
|
| 36 |
-
| Architecture | Conformer-CTC, Metro-style Macaron blocks |
|
| 37 |
-
| Size | 57.9M inference parameters, d=384, 12 layers, 6 heads |
|
| 38 |
-
| Languages | Urdu (id 1), Sindhi (id 2). Id 0 is unconditioned |
|
| 39 |
-
| Conditioning | language embedding and a language-ID head |
|
| 40 |
-
| Tokenizer | character, `char_urd_snd` |
|
| 41 |
-
| Frontend | 16 kHz, 80-bin log-mel, 4× conv subsampling, 25 frames/s |
|
| 42 |
-
| Objective | `(1 − 0.3) · CTC + 0.3 · auxiliary CTC` at layer 6 |
|
| 43 |
-
| Decoder in this table | greedy CTC, no n-gram |
|
| 44 |
-
| Training | 40,000 steps, bf16, RTX 3090, about 9,114 hours of audio seen, ~50/50 Urdu/Sindhi |
|
| 45 |
-
| Code | Apache-2.0 |
|
| 46 |
-
|
| 47 |
|
| 48 |
## Use
|
| 49 |
|
|
|
|
| 19 |
|
| 20 |
Canopy-M was trained on Urdu and Sindhi. The numbers below are the evaluation.
|
| 21 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
|
| 23 |
|
| 24 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
|
| 26 |
## Use
|
| 27 |
|