File size: 1,959 Bytes
ca57e75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9d14cb7
 
 
 
 
 
 
ca57e75
 
 
 
 
 
 
 
 
02947d1
 
 
 
 
 
ca57e75
 
02947d1
ca57e75
 
02947d1
ca57e75
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
---
license: apache-2.0
base_model: openai/whisper-large-v2
library_name: basert
pipeline_tag: automatic-speech-recognition
tags:
  - basert
  - apple-silicon
  - whisper
  - speech-recognition
---

# whisper-large-v2

[BaseRT](https://github.com/basecompute/baseRT) `.base` build of
[`openai/whisper-large-v2`](https://huggingface.co/openai/whisper-large-v2),
OpenAI's 1.55B-parameter Whisper speech-recognition model,
for fast local transcription on Apple Silicon.

## Files

| File | Precision | Size |
|------|-----------|------|
| `whisper-large-v2-F16.base` | float16 | 3.10 GB |
| `whisper-large-v2-Q8.base` | 8-bit linears, f16 embeddings/conv/norms | 1.68 GB |
| `whisper-large-v2-Q4.base` | 4-bit linears, f16 embeddings/conv/norms | 987 MB |

F16 and Q8 are transcription-quality equivalent (Q8 passes the same
word-error parity gates against reference openai-whisper). Q4 is the
smallest and remains accurate; on some smaller variants it can
occasionally repeat a word in timestamped beam decoding.

## Usage

```bash
curl -LsSf https://basecompute.co/install.sh | sh
basert serve --model whisper-large-v2-F16.base
```

`POST /v1/audio/transcriptions` (multipart or JSON) returns `json`, `text`,
`srt`, `vtt`, or `verbose_json` (with per-segment `avg_logprob` /
`no_speech_prob` / `compression_ratio` / `temperature` and the detected
`language`), with optional SSE streaming and `POST /v1/audio/translations`. Supported
request fields: `language` (or `"auto"` to detect), `prompt`
(initial prompt / vocabulary bias), `task` (`transcribe`/`translate`). Or transcribe directly
from the CLI:

```bash
basert-transcribe whisper-large-v2-F16.base audio.wav --lang auto
```

Multilingual: pass `language` (e.g. `--lang de`), or `auto` to detect, across the 99 supported languages. `task=translate` produces English from any source language (also via `POST /v1/audio/translations`).

Released under the apache-2.0 license, inherited from the base model.