Downloads · 30 days
22
2% of all-time downloads
alekshandru/qwen3-asr-0.6b-onnx
qwen3-asr-0.6b-onnx is a automatic speech recognition model from alekshandru. Use it when you need speech turned into text. It is set up for onnxruntime. The card lists the license as apache-2.0.
ONNX export of Qwen/Qwen3-ASR-0.6B for use with ONNX Runtime.
Downloads · 30 days
22
2% of all-time downloads
All-time downloads
1.4K
Public
Repo size
44.4 GB
Likes
0
Public
Click a slice to open those files.
.data4 GB · 43%
From the Hugging Face model README
ONNX export of Qwen/Qwen3-ASR-0.6B for use with ONNX Runtime.
Includes FP32 and int4 MatMulNBits variants. Quantized files use .int4. naming and coexist with FP32 in the same directory. The transcribe-rs engine auto-detects quantized models by file presence.
| File | Description |
|---|---|
encoder.onnx | Audio encoder — mel spectrogram to features (weights inlined) |
decoder_init.onnx | Decoder prefill — accepts input_ids, outputs logits + KV cache |
decoder_step.onnx | Decoder autoregressive step — accepts input_embeds + KV cache |
decoder_weights.data | Shared external weights for both FP32 decoders (loaded once) |
| File | Description |
|---|---|
encoder.int4.onnx | Encoder (FP32 weights — same as encoder.onnx) |
decoder_init.int4.onnx | int4 decoder prefill |
decoder_step.int4.onnx | int4 decoder step |
decoder_weights.int4.data | Shared external weights for both int4 decoders (loaded once) |
| File | Description |
|---|---|
embed_tokens.bin | Token embedding matrix [151936, 1024], float16 |
tokenizer.json | HuggingFace tokenizer |
config.json | Architecture config, special tokens, mel params |
preprocessor_config.json | Mel spectrogram parameters (WhisperFeatureExtractor format) |
The two decoder models use different input strategies:
decoder_init (prefill) accepts input_ids and has the embedding table in its graph — handles audio feature scatter internallydecoder_step (autoregressive) accepts pre-looked-up input_embeds — keeps the embedding table out of its graphThe consumer loads embed_tokens.bin once at startup and performs the embedding lookup per token before calling decoder_step.
Identical to Whisper: 16kHz, 128 bins, n_fft=400, hop_length=160, Hann window, Slaney mel scale, 0-8kHz.
Also documented in preprocessor_config.json (WhisperFeatureExtractor format).
encoder.onnx (or encoder.int4.onnx): mel → audio_features<|im_start|>system<|im_end|><|im_start|>user<|audio_start|><|audio_pad|>...<|audio_end|><|im_end|><|im_start|>assistantdecoder_init.onnx (or .int4): input_ids + audio_features + audio_offset → logits + KV cachedecoder_step.onnx (or .int4): look up embed_tokens.bin → input_embeds + KV cache → logits, until EOS| Token | ID |
|---|---|
<|audio_start|> | 151669 |
<|audio_end|> | 151670 |
<|audio_pad|> | 151676 |
<|im_end|> (EOS) | 151645 |
<|endoftext|> (EOS) | 151643 |
Exported with qwen3-asr-onnx.