{
  "schema_version": 1,
  "_comment": "Whisper-small ASR (encoder-decoder). Host log-mel + host conv stem (the conv is HTP-hostile in-graph) feed the encoder graph (input encoder_features); the autoregressive decoder (causal self-attention + cross-attention to the cached encoder states, learned positional embeddings, tied lm-head) emits tokens, detok'd on the host. Gated on WER vs HF. Same proven asr_transcribe host-op path as whisper-base (12 layers / d768 / 12 heads).",
  "model": {
    "name": "whisper-small",
    "family": "asr",
    "dsp_arch": "v81",
    "tokenizer_pre": "whisper"
  },
  "params": {
    "hidden": 768,
    "vocab": 51865,
    "n_layers": 12,
    "max_ctx": 200,
    "kv_dim": 768,
    "head_dim": 64,
    "rope_theta": 10000.0,
    "eos_token_id": 50257
  },
  "artifacts": {
    "contexts": {
      "encoder": {
        "bin": "whispersmall_enc_f16.bin"
      },
      "decoder": {
        "bin": "whispersmall_dec_f16.bin"
      }
    },
    "tokenizer": "tokenizer.json"
  },
  "plan": {
    "steps": [
      {
        "host": "asr_transcribe",
        "params": {
          "encoder": "whispersmall_enc_f16",
          "decoder": "whispersmall_dec_f16",
          "sot": 50258,
          "mel_filters": "whisper_small_mel_filters.bin",
          "conv_weights": "whisper_conv_stem.bin",
          "n_mels": 80
        }
      }
    ]
  }
}