justtalk-coreml — CoreML speech models for Just Talk

CoreML (.mlmodelc) conversions of permissively licensed Whisper models, used on-device by Just Talk via WhisperKit. We host them ourselves so the exact weights we tested stay pinned and reachable, independent of any third-party repo.

Everything here is a derivative work of an upstream Apache-2.0 model. Nothing in this repository is trained by us.

Contents

Directory	Upstream	Upstream licence	What we changed
openai_whisper-small/	openai/whisper-small	Apache-2.0 (model card); the source repo ships MIT	Converted PyTorch → CoreML .mlmodelc (MelSpectrogram, AudioEncoder, TextDecoder), fp16, ANE-oriented attention layout
vasista22_whisper-hindi-medium/	vasista22/whisper-hindi-medium	Apache-2.0	Same conversion. Alignment heads borrowed from openai/whisper-medium where the fine-tune omitted them
vasista22_whisper-hindi-medium_q8/	same as above	Apache-2.0	Same conversion, plus 8-bit per-grouped-channel palettization of the encoder and decoder

A 4-bit variant was produced during evaluation and is deliberately not published: it measurably degraded Hindi accuracy (~27% WER vs ~18% for 8-bit).

Provenance

• Conversion tool: argmaxinc/whisperkittools, pinned at commit 84f77a8 (MIT, Copyright (c) 2024 Argmax, Inc.).
• Conversion performed 2026-07-15 by the script dev/convert-indic-whisper.sh in the Just Talk repo.
• Upstream revisions: the exact source revision was not recorded at conversion time. For reference, as of 2026-09-16 the upstream repositories are at openai/whisper-small → 973afd24965f (last modified 2024-02-29) and vasista22/whisper-hindi-medium → d53532a4dc1d (last modified 2023-04-24). Both have been unchanged since well before our conversion date, but we state this as a reference point rather than a claim about the bytes we converted.

Licence

Apache License 2.0 — full text in LICENSE; attribution and modification notices in NOTICE. Redistributed by Trika. If you use these weights, keep both files with them.