Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -22,6 +22,13 @@ for fast local transcription on Apple Silicon.
|
|
| 22 |
| File | Precision | Size |
|
| 23 |
|------|-----------|------|
|
| 24 |
| `whisper-large-v2-F16.base` | float16 | 3.10 GB |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
|
| 26 |
## Usage
|
| 27 |
|
|
|
|
| 22 |
| File | Precision | Size |
|
| 23 |
|------|-----------|------|
|
| 24 |
| `whisper-large-v2-F16.base` | float16 | 3.10 GB |
|
| 25 |
+
| `whisper-large-v2-Q8.base` | 8-bit linears, f16 embeddings/conv/norms | 1.68 GB |
|
| 26 |
+
| `whisper-large-v2-Q4.base` | 4-bit linears, f16 embeddings/conv/norms | 987 MB |
|
| 27 |
+
|
| 28 |
+
F16 and Q8 are transcription-quality equivalent (Q8 passes the same
|
| 29 |
+
word-error parity gates against reference openai-whisper). Q4 is the
|
| 30 |
+
smallest and remains accurate; on some smaller variants it can
|
| 31 |
+
occasionally repeat a word in timestamped beam decoding.
|
| 32 |
|
| 33 |
## Usage
|
| 34 |
|