add fireredasr2-aed stt model

[mm65x]

Context

FireRedASR2-AED (https://github.com/FireRedTeam/FireRedASR2S) from Xiaohongshu/AntGroup. conformer encoder + transformer decoder model for Chinese and English ASR. ~1.18B params, strong on Mandarin (including 20+ dialects/accents), English, and code-switching.

Description

adds FireRedASR2-AED to the stt pipeline. the model uses a 16 layer conformer encoder with relative positional attention (Transformer-XL style with pos_bias_u/v) and a 16 layer transformer decoder with beam search decoding.

the upstream repo distributes weights as a single model.pth.tar with an argparse namespace for config, so this includes a conversion script that extracts weights to safetensors, converts the kaldi CMVN stats to json, and copies the tokenizer files.

the tokenizer is a hybrid system - Chinese characters are treated as individual tokens while English text goes through SentencePiece BPE (~1000 merges).

Changes in the codebase

• mlx_audio/stt/models/fireredasr2/fireredasr2.py - model implementation (fbank frontend, CMVN, Conv2d subsampling, conformer encoder, transformer decoder, beam search, generate)
• mlx_audio/stt/models/fireredasr2/config.py - config dataclasses
• mlx_audio/stt/models/fireredasr2/convert.py - pth.tar → safetensors converter
• mlx_audio/stt/models/fireredasr2/README.md - setup + usage
• mlx_audio/stt/models/__init__.py, mlx_audio/stt/utils.py - registration

Changes outside the codebase

none.

Additional information

• verified end-to-end against the PyTorch reference implementation on Chinese, English, and out-of-distribution (German) audio - transcriptions match exactly
• encoder intermediate outputs compared numerically: mean abs diff of 0.005 across 16 conformer blocks (expected for float32 cross-framework differences)
• beam search works for beam_size 1, 3, 5, 10
• users need to download weights from ModelScope and run the conversion script (instructions in the README)

Checklist

• Tests added/updated
• Documentation updated
• Issue referenced - closes Support FireRedASR2S #555

[lucasnewman]

@mm65x Can you just upload the converted model to huggingface? Then we don't need the kaldiio dependency.

[mm65x]

@lucasnewman uploaded the converted model here: https://huggingface.co/mlx-community/FireRedASR2-AED-mlx

i've also removed convert.py, updated the README to use mlx-community/FireRedASR2-AED-mlx, and fixed the test file formatting in the latest commit. let me know if there's anything else needed here.

[lucasnewman]

Awesome, thanks!