aoiandroid/Qwen3-ASR-1.7B-CoreML

Qwen3-ASR-1.7B CoreML

CoreML conversion of Qwen/Qwen3-ASR-1.7B for on-device speech-to-text on macOS/iOS.

Optimized with FP32 compute + ANEMLL RMSNorm + mixed-precision INT8 to prevent decoder overflow while keeping model size small.

Files

File	Description	Size
qwen3_asr_encoder_int8.mlpackage/	Audio encoder (Conv2D + 24-layer Transformer + projection). FP16 compute, INT8 per-channel weights.	304 MB
qwen3_asr_decoder_f32_anemll_int8-mixed.mlpackage/	Decoder (28-layer causal LLM + LM head). FP32 compute, ANEMLL RMSNorm, mixed-precision INT8.	2.8 GB
qwen3_asr_embeddings.bin	Token embedding table (FP16, shape 151936 x 2048). Loaded separately to avoid quantizing embeddings.	594 MB

Architecture

Audio (16kHz) → Mel Spectrogram (128 bins)
  → Encoder (100-frame chunks → 13 tokens each)
  → Embedding lookup + prompt construction
  → Decoder (autoregressive, RoPE as inputs, KV cache)
  → Token IDs → Text

Encoder

• 3x Conv2D (stride 2) + 24-layer Transformer + 2-layer projection
• Input: [1, 1, 128, 100] mel chunk → Output: [1, 13, 2048] audio features
• Bidirectional attention, sinusoidal positional embeddings (reset per chunk)
• INT8 per-channel quantization (first conv layer kept at FP16)

Decoder

• 28-layer causal Transformer with GQA (16 Q heads, 8 KV heads, head_dim=128)
• SwiGLU MLP (intermediate_size=11008)
• RoPE theta=1,000,000 (NeoX/split-half style, passed as cos/sin inputs)
• ANEMLL-style RMSNorm: LayerNorm([x, -x]) trick for numerical stability
• Mixed-precision INT8: first/last layer, norms, LM head kept at FP16

Key Quality Improvements

1. FP32 compute for decoder — Hidden states reach 10,876 at layer 25; x^2 = 118M >> FP16 max 65,504. FP32 is mandatory.
2. ANEMLL RMSNorm — Uses native layer_norm op via [x, -x] concatenation for GPU/ANE precision.
3. Mixed-precision INT8 — Skips 72 sensitive ops (embeddings, first/last layer, all norms, LM head).
4. RoPE as inputs — Precomputed sin/cos passed in to avoid in-model precision loss.
5. Embeddings extracted — 151,936-token embedding table stored as raw FP16 binary, not quantized.

Quality Benchmarks

English (JFK speech, 11s)

Pipeline	Transcription	Match
PyTorch FP32	"And so, my fellow Americans, ask not what your country can do for you. Ask what you can do for your country."	Reference
CoreML INT8	Same text (minor punctuation difference: "." vs ";")	MATCH

Taiwan Mandarin (Taiwan-Tongues test set, 20 samples, 60s total)

Metric	Value
PyTorch vs CoreML match rate	19/20 (95%)
Average CER (PyTorch)	0.2009
Average CER (CoreML INT8)	0.2064
CER difference	+0.0056

Usage Notes

• macOS 14+ required (minimum deployment target)
• Decoder runs on GPU only (FP32 compute prevents ANE execution)
• KV cache must be initialized with 1 dummy slot (CoreML crashes on size-0 dynamic dims); mask the dummy position with -inf
• Encoder processes mel in 100-frame chunks — split mel, run encoder per chunk, concatenate outputs

Conversion Details

Converted using coremltools 9.0 with PyTorch 2.10.0. Conversion scripts available at the source repository.

compute_precision: FLOAT32 (decoder), FLOAT16 (encoder)
minimum_deployment_target: macOS14
quantization: linear_symmetric INT8 per-channel (mixed-precision for decoder)

License

Apache 2.0 (same as base model)