arXiv · 2022 · Preprint
Gao et al. (Alibaba Group) · → Paper · Demo: ? · Code: ✓
Paraformer is a single-step non-autoregressive ASR model that combines a CIF-based predictor for token-length estimation with a glancing language model sampler to recover context interdependence, achieving accuracy comparable to autoregressive transformers with more than 10x inference speedup on Mandarin benchmarks up to 20,000 hours.
Citation Stub
This paper is not a speech generation paper but is cited by the corpus. See Context in Speech Generation below for why it is relevant.
Context in Speech Generation
Paraformer introduces a parallel Transformer ASR architecture that eliminates the sequential decoding bottleneck of standard attention-encoder-decoder models. The key insight is pairing a continuous integrate-and-fire (CIF) predictor — which estimates output token count and produces acoustic embeddings — with a glancing language model (GLM) sampler that injects target-side context during training to reduce substitution errors caused by the independence assumption in single-step non-autoregressive decoding. On AISHELL-1 the system achieves 5.2% CER, matching its autoregressive baseline, while running at an RTF roughly 12 times lower. TTS and speech LM pipelines that require a fast, high-accuracy ASR evaluator can use Paraformer as an intelligibility evaluator in place of slower autoregressive ASR systems, with inference latency that is compatible with large-scale automatic evaluation loops.
Wiki Connections
- 1808.10583 — AISHELL-2 dataset used as a benchmark in Paraformer’s evaluation