🔄 [处理中][wt-audio8-asr] feat: Add Audio8-ASR-0.1B community model port (audio8_asr)

[gqf2008]

Goal

Add native audio.cpp support for Audio8/Audio8-ASR-0.1B (HF: Audio8/Audio8-ASR-0.1B) as community family audio8_asr.

Audio8-ASR-0.1B is a compact multilingual ASR model (en/zh/yue/ja/ko/fr/de): Qwen3-ASR-0.6B audio encoder + 4-layer MLP tower adapter + 8-layer Qwen2-style decoder (hidden 512, tied embeddings, vocab 151936). 324M params total, 648 MB bf16. Self-reported Open ASR Leaderboard 7-split mean WER 7.03% (EN), WenetSpeech net CER 7.98 (ZH). Offline, ≤30 s, greedy.

Scope

• Python reference environment + golden parity outputs (torch CPU, transformers, trust_remote_code)
• GGUF converter tools/community_models/convert_audio8_asr.py (safetensors → GGUF, metadata-embedded configs)
• C++ community family src/community_models/audio8_asr/ (loader/assets/session/prompt/tower adapter/decoder)
• Reuse qwen3_asr audio encoder + Whisper 128-mel frontend (configs are bit-identical)
• Decoder via shared QwenCausalDecoder with use_qk_norm=false (Qwen2 style), tied lm_head
• model_specs/audio8_asr.json — offline-only modes, local-conversion-only download (CC-BY-NC-4.0, follow mms_forced_aligner precedent: no redistribution)
• Registration: CMakeLists audiocpp_add_model, loader catalog sync check, README supported-model table, docs/community_models/audio8_asr.md
• Unit tests + parity evidence vs Python reference

Out of scope (phase 1)

• Hotword logit boosting (needs decode-time logit-bias hook; later phase)
• Word timestamps, streaming mode
• GGUF redistribution (license is CC-BY-NC-4.0 — conversion is local-only)

Acceptance criteria

• audiocpp_cli --list-loaders --json includes audio8_asr; python3 tools/check_loader_catalog_sync.py passes
• Converted GGUF loads and transcribes a sample wav; output matches Python reference (greedy, 30 s window) exactly or with documented tolerance
• Build + unit tests green locally (macOS, CPU/Metal backend)
• Parity/build/run commands documented in the PR

References

• HF model card: https://huggingface.co/Audio8/Audio8-ASR-0.1B (CC-BY-NC-4.0)
• Architecture notes: whisper 128-mel frontend → Qwen3-ASR encoder (896/18L/14H/ffn3584) → 4× pre-LN residual MLP tower (1024→4096, GELU) → avg-pool merge_factor 4 → LN+Linear(1024→512) → set_rows at <|audio|> (151646); prompt <|user|><|begin_of_audio|><|audio|>×N<|end_of_audio|>…<|assistant|>, N=((samples/160+1)//2)//4
• Related but non-overlapping: PR feat: Supporting Audio8_TTS models #333 (audio8_tts, TTS only)
• NC-license precedent: model_specs/mms_forced_aligner.json (local-only conversion)

[gqf2008]

Progress update (PR: #337):

• Python reference environment + golden parity outputs
• GGUF converter tools/community_models/convert_audio8_asr.py
• C++ community family src/community_models/audio8_asr/
• Reuses qwen3_asr audio encoder + Whisper 128-mel frontend via a renaming tensor source (no core changes)
• Decoder via shared QwenCausalDecoder with use_qk_norm=false, attention biases, tied lm_head
• model_specs/audio8_asr.json — offline-only, CC-BY-NC local-conversion-only (mms_forced_aligner precedent)
• Registration: CMake audiocpp_add_model (DEPENDS qwen3_asr), loader catalog sync green, README + community docs, webui catalog entry
• Tests: test_audio8_asr_units (formula + bf16 rounding) and test_audio8_asr_golden_transcription (end-to-end); exact-match parity vs fp32 reference on Metal + CPU with the Q8_0 GGUF, plus F16 GGUF and direct safetensors loading
• Independent code review (running)