Add FFASR leaderboard result (#1)
Browse files
README.md
CHANGED
|
@@ -18,12 +18,26 @@ Tiro is a far-field fine-tune of
|
|
| 18 |
performance on reverberant and noisy speech, scaling as conditions get harder, while
|
| 19 |
maintaining general ASR capabilities. See benchmarks below for more details.
|
| 20 |
|
|
|
|
|
|
|
| 21 |
English only, bf16. Same architecture and parameter count as the base.
|
| 22 |
|
| 23 |
## Benchmarks
|
| 24 |
|
| 25 |
Lower is better. ✅ marks the better result.
|
| 26 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
| Benchmark | Condition | **Base (WER)** | **Tiro (WER)** |
|
| 28 |
|---|---|---:|---:|
|
| 29 |
| [VOiCES](https://iqtlabs.github.io/voices/) | real rooms, replayed speech | 8.88 | **8.30** ✅ |
|
|
|
|
| 18 |
performance on reverberant and noisy speech, scaling as conditions get harder, while
|
| 19 |
maintaining general ASR capabilities. See benchmarks below for more details.
|
| 20 |
|
| 21 |
+
Currently **#1 on the [FFASR leaderboard](https://huggingface.co/spaces/treble-technologies/ffasr)**.
|
| 22 |
+
|
| 23 |
English only, bf16. Same architecture and parameter count as the base.
|
| 24 |
|
| 25 |
## Benchmarks
|
| 26 |
|
| 27 |
Lower is better. ✅ marks the better result.
|
| 28 |
|
| 29 |
+
### [FFASR Leaderboard](https://huggingface.co/spaces/treble-technologies/ffasr)
|
| 30 |
+
|
| 31 |
+
| Condition | **Base (WER)** | **Tiro (WER)** |
|
| 32 |
+
|---|---:|---:|
|
| 33 |
+
| **Average** | 13.41 | **12.01** ✅ |
|
| 34 |
+
| Near field speech | **3.76** ✅ | 3.80 |
|
| 35 |
+
| High SNR | 6.74 | **6.22** ✅ |
|
| 36 |
+
| Mid SNR | 13.89 | **12.00** ✅ |
|
| 37 |
+
| Low SNR | 29.26 | **26.02** ✅ |
|
| 38 |
+
|
| 39 |
+
### Other benchmarks
|
| 40 |
+
|
| 41 |
| Benchmark | Condition | **Base (WER)** | **Tiro (WER)** |
|
| 42 |
|---|---|---:|---:|
|
| 43 |
| [VOiCES](https://iqtlabs.github.io/voices/) | real rooms, replayed speech | 8.88 | **8.30** ✅ |
|