Add Doppel model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: coreml
|
| 4 |
+
pipeline_tag: automatic-speech-recognition
|
| 5 |
+
base_model: Qwen/Qwen3-ASR-0.6B-hf
|
| 6 |
+
inference: false
|
| 7 |
+
tags:
|
| 8 |
+
- coreml
|
| 9 |
+
- apple-silicon
|
| 10 |
+
- macos
|
| 11 |
+
- on-device
|
| 12 |
+
- doppel
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# Doppel ASR companion
|
| 16 |
+
|
| 17 |
+
This repository is Doppel's distribution channel for an app-specific Core ML
|
| 18 |
+
conversion of [`Qwen/Qwen3-ASR-0.6B-hf`](https://huggingface.co/Qwen/Qwen3-ASR-0.6B-hf).
|
| 19 |
+
It exists so supported Doppel releases can perform speech recognition on the
|
| 20 |
+
user's Mac.
|
| 21 |
+
|
| 22 |
+
## Intended use
|
| 23 |
+
|
| 24 |
+
These artifacts are intended to be downloaded, verified, and invoked by
|
| 25 |
+
Doppel on Apple-silicon Macs running macOS 15 or later. Inference stays on the
|
| 26 |
+
device.
|
| 27 |
+
|
| 28 |
+
This is not a standalone ASR distribution or a supported developer API. The
|
| 29 |
+
repository does not include Doppel's audio preprocessing, tokenizer assets,
|
| 30 |
+
runtime orchestration, output handling, or compatibility logic. It is not
|
| 31 |
+
expected to work with generic Transformers, Core ML sample code, or Hugging
|
| 32 |
+
Face hosted inference. For general-purpose use, start with the upstream model.
|
| 33 |
+
|
| 34 |
+
## Artifact and integrity notes
|
| 35 |
+
|
| 36 |
+
The compiled Core ML components are modified from the upstream checkpoint for
|
| 37 |
+
Doppel's fixed runtime contract. `manifest.json` is the authoritative artifact
|
| 38 |
+
inventory and records the integrity hashes checked by the app.
|
| 39 |
+
|
| 40 |
+
The repository contains no user audio, transcripts, prompts, indexes,
|
| 41 |
+
telemetry, or other customer data.
|
| 42 |
+
|
| 43 |
+
## Limitations
|
| 44 |
+
|
| 45 |
+
Speech recognition can be inaccurate, especially with noise, overlapping
|
| 46 |
+
speakers, uncommon names, code switching, or domain-specific language. Doppel
|
| 47 |
+
must not treat a transcript as a verbatim or safety-critical record without
|
| 48 |
+
human review.
|
| 49 |
+
|
| 50 |
+
## License and attribution
|
| 51 |
+
|
| 52 |
+
The upstream model is provided under the
|
| 53 |
+
[Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).
|
| 54 |
+
This converted distribution remains subject to the applicable upstream terms.
|
| 55 |
+
Doppel provides no support or warranty for use outside the Doppel app.
|