README.md

---
license: mit
tags:
  - automatic-speech-recognition
  - audio
  - whisper
  - ferrotorch
---

# `ferrotorch/whisper-tiny-encoder`

whisper-tiny encoder (openai/whisper-tiny). 4-layer 6-head Transformer audio encoder, d_model=384, encoder_ffn_dim=1536, num_mel_bins=80, max_source_positions=1500. MIT-licensed. Pinned encoder-only — decoder/proj_out weights are dropped from this mirror. Real-artifact baseline for audio encoder parity vs transformers (#1149).

## Provenance

* Upstream: `openai/whisper-tiny` (mit).
* Conversion script:
  [`ferrotorch/scripts/pin_pretrained_whisper_weights.py`](https://github.com/dollspace-gay/ferrotorch/blob/main/scripts/pin_pretrained_whisper_weights.py).
* Ferrotorch issue: <https://github.com/dollspace-gay/ferrotorch/issues/1149>.
* SHA-256 of `model.safetensors` (this file is pinned in
  `ferrotorch-hub/src/registry.rs`): `4ce29194b87ef05385203f8b09914f5c3b060200c2b503d6d420459ffb80a294`.
* Number of trainable parameters in the encoder slice:
  **8,208,384**.
* Config snapshot: d_model=384,
  encoder_layers=4,
  encoder_attention_heads=6,
  encoder_ffn_dim=1536,
  num_mel_bins=80,
  max_source_positions=1500,
  activation_function='gelu'.
* Non-encoder keys dropped from the upstream checkpoint (this
  mirror is encoder-only): 100 total, first few:
  `['model.decoder.embed_positions.weight', 'model.decoder.embed_tokens.weight', 'model.decoder.layer_norm.bias']`.

## Value-parity probe

Three extra files are uploaded so the ferrotorch-side harness can
reproduce the parity verdict without re-running the upstream
Whisper model:

* `_value_parity_audio.bin` — deterministic synthetic 30-second
  audio (sum of three sine waves with a slow envelope),
  16 kHz mono float32, shape `[1, 480000]`.
* `_value_parity_mel.bin` — `WhisperFeatureExtractor(audio)`
  output `[1, 80, 3000]` float32 from the upstream feature
  extractor. The Rust-side `ferrotorch_whisper::audio` is
  compared against this.
* `_value_parity_encoder_output.bin` — float32 encoder hidden
  states `[1, 1500, 384]` from
  `WhisperModel.encoder(input_features=mel).last_hidden_state`.
  Format: `[u32 ndim][u32 × ndim shape][f32 × prod(shape)]`
  little-endian (matches every other ferrotorch dump).

## How to load

```rust
use ferrotorch_whisper::{
    HfWhisperConfig, WhisperConfig, load_whisper_encoder,
};
use ferrotorch_hub::{HubCache, hf_download_model};

let cache = HubCache::with_default_dir();
let repo_dir = hf_download_model("ferrotorch/whisper-tiny-encoder", "main", &cache)?;
let hf_cfg = HfWhisperConfig::from_file(repo_dir.join("config.json"))?;
let cfg = WhisperConfig::from_hf(&hf_cfg)?;
let (encoder, _drop_report) = load_whisper_encoder::<f32>(
    &repo_dir.join("model.safetensors"),
    cfg,
    /* strict = */ false,
)?;
```

## Upstream license

```
MIT License

Copyright (c) 2022 OpenAI

Permission is hereby granted, free of charge, to any person obtaining a
copy of this software and associated documentation files (the
"Software"), to deal in the Software without restriction, including
without limitation the rights to use, copy, modify, merge, publish,
distribute, sublicense, and/or sell copies of the Software, and to
permit persons to whom the Software is furnished to do so, subject to
the following conditions:

The above copyright notice and this permission notice shall be included
in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

```