Qwen3-ASR-1.7B

Converts audio waveforms to text across multiple languages using a transformer-based encoder-decoder architecture optimized for 1.7B parameters. The model processes raw audio through a mel-spectrogram frontend, encodes acoustic features via a conformer-style encoder, and decodes to text tokens via an autoregressive decoder. Supports streaming and batch inference modes with dynamic quantization for edge deployment.

Processes audio in real-time chunks (typically 320-640ms windows) using a streaming-compatible encoder-decoder that maintains hidden state across chunks, enabling sub-second latency transcription without buffering entire audio files. Implements a sliding window attention mechanism in the encoder to avoid reprocessing overlapping audio frames, and uses incremental decoding to emit partial hypotheses as new audio arrives.

Supports dynamic quantization (INT8/FP16) and static quantization (INT4/INT8) via ONNX Runtime and TensorRT, reducing model size from 1.7B parameters (~3.4GB in FP32) to 850MB-1.7GB depending on quantization scheme. Quantization is applied post-training without retraining, preserving accuracy within 1-3% of the original model while reducing memory footprint and inference latency by 2-4x on CPU and 1.5-2x on GPU.

Supports parameter-efficient fine-tuning via LoRA (Low-Rank Adaptation) and full fine-tuning on custom speech datasets. The model's encoder and decoder can be selectively frozen, allowing adaptation of only the attention layers or decoder to new acoustic domains (e.g., medical terminology, accent-specific speech). Fine-tuning uses CTC loss for the encoder and cross-entropy loss for the decoder, with support for mixed-precision training (FP16/BF16) to reduce memory requirements.

Outputs per-token confidence scores derived from the decoder's softmax probabilities, enabling downstream applications to identify low-confidence regions in transcripts. The model also supports beam search decoding (beam width 1-5) to generate multiple hypothesis transcripts with associated log-probabilities, allowing uncertainty quantification via hypothesis diversity and score margins. Confidence scores can be aggregated at word or utterance level for downstream filtering or rejection.

Handles code-switching (mixing multiple languages within a single utterance) by training on multilingual data with language-agnostic acoustic features and a shared vocabulary across languages. The model does not require explicit language tags at inference time; instead, it learns to recognize language boundaries implicitly through acoustic and linguistic context. Supports seamless transcription of utterances like 'Hello, 你好, bonjour' without language-specific preprocessing.

Generates word-level and sub-word-level timestamps by aligning the decoder's output tokens with the encoder's frame-level acoustic features. Uses a forced alignment algorithm (CTC alignment or attention-based alignment) to map each output token to its corresponding time range in the input audio. Timestamps are returned as start/end times in milliseconds, enabling precise synchronization with video or other time-indexed media.

Supports efficient batch inference by dynamically grouping audio samples of varying lengths into batches, padding shorter sequences and masking padded regions to avoid unnecessary computation. Uses a bucketing strategy to group similar-length audios together, reducing padding overhead. Batch processing is optimized for both GPU (via CUDA kernels) and CPU (via vectorized operations), with configurable batch sizes and sequence length limits.