Card: parameter breakdown, accuracy, input, file sizes, speed and memory at a glance
Browse files
README.md
CHANGED
|
@@ -56,11 +56,16 @@ Streaming Thai speech-to-text with 0.64B parameters. It writes Thai speech down
|
|
| 56 |
| | |
|
| 57 |
|---|---|
|
| 58 |
| Architecture | Cache-aware streaming FastConformer encoder + RNN-Transducer decoder, prompt-conditioned (NeMo `EncDecRNNTBPEModelWithPrompt`) |
|
| 59 |
-
| Parameters | 640.4 million |
|
|
|
|
| 60 |
| Latency settings | 80, 320, 560 and 1120 ms (`att_context_size` `[56,0]`, `[56,3]`, `[56,6]`, `[56,13]`), chosen at decode time |
|
| 61 |
| Language prompts | `th-TH`: everything in Thai script. `th-mix`: English words in Latin letters |
|
| 62 |
| Tokenizer | 14,988 pieces: NVIDIA's 13,087 multilingual pieces and 1,901 new Thai pieces |
|
| 63 |
| Training data | 12,373 hours of Thai speech and 961 hours of English (LibriSpeech) |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 64 |
| Licence | OpenMDW-1.1 |
|
| 65 |
|
| 66 |
## Results
|
|
|
|
| 56 |
| | |
|
| 57 |
|---|---|
|
| 58 |
| Architecture | Cache-aware streaming FastConformer encoder + RNN-Transducer decoder, prompt-conditioned (NeMo `EncDecRNNTBPEModelWithPrompt`) |
|
| 59 |
+
| Parameters | 640.4 million (0.64B): encoder 609.1 million (24 FastConformer layers, 1,024 wide), language-prompt layer 4.5 million, RNN-T decoder 16.2 million (2-layer LSTM of 640) and joint network 10.7 million |
|
| 60 |
+
| Accuracy, 560 ms | % of characters wrong: FLEURS 8.47, GigaSpeech2 15.35, TVSpeech 12.41, LOTUSDIS 20.77, Isan 34.90 (average 18.38; see [Results](#results)) |
|
| 61 |
| Latency settings | 80, 320, 560 and 1120 ms (`att_context_size` `[56,0]`, `[56,3]`, `[56,6]`, `[56,13]`), chosen at decode time |
|
| 62 |
| Language prompts | `th-TH`: everything in Thai script. `th-mix`: English words in Latin letters |
|
| 63 |
| Tokenizer | 14,988 pieces: NVIDIA's 13,087 multilingual pieces and 1,901 new Thai pieces |
|
| 64 |
| Training data | 12,373 hours of Thai speech and 961 hours of English (LibriSpeech) |
|
| 65 |
+
| Input | 16 kHz mono audio (128 log-mel bands); any length when cut at pauses, as the included server does |
|
| 66 |
+
| Files | Two `.nemo` files of 2.56 GB each, 32-bit weights (see [Use](#use)). An ONNX export, full precision and int8, is at [iapp/OpenThai2.0-ASR-Streaming-ONNX](https://huggingface.co/iapp/OpenThai2.0-ASR-Streaming-ONNX) |
|
| 67 |
+
| Speed | One NVIDIA H100: files about 257× faster than real time, about 20 live streams at once at 560 ms. A CPU with 8 threads: files about 4× real time, a live stream keeps up at 560 and 320 ms |
|
| 68 |
+
| Memory | About 2.6 GB for the weights; about 6.5 GB of RAM in all on a CPU |
|
| 69 |
| Licence | OpenMDW-1.1 |
|
| 70 |
|
| 71 |
## Results
|