Researchers propose a Deep Feedforward Sequential Memory Network (DFSMN) for speech synthesis that matches BLSTM quality while using only a quarter of the model size and achieving four times faster inference, making it ideal for memory‑constrained, real‑time edge devices.
This article introduces the Deep‑FSMN (DFSMN) architecture and its integration with low‑frame‑rate (LFR) processing, showing how the combined LFR‑DFSMN acoustic model achieves higher accuracy, smaller model size, faster training, and lower latency than traditional BLSTM‑based speech recognition systems on both English and Chinese large‑vocabulary tasks.
This paper introduces a Deep Feedforward Sequential Memory Network (DFSMN) for statistical parametric speech synthesis that matches BLSTM quality with only a quarter of the model size and four times faster inference, making it ideal for memory‑constrained, real‑time IoT devices.
Alibaba’s DAMO Academy unveiled the DFSMN speech‑recognition model, open‑sourced on GitHub, which sets a new 96.04% accuracy record on LibriSpeech, trains three times faster than LSTM, and powers real‑world demos like AI cashiers and metro ticket machines.
Alibaba's DAMO Academy has open‑sourced the Deep‑Feedforward Sequential Memory Network (DFSMN), a next‑generation speech‑recognition model that achieves a world‑record 96.04% accuracy on LibriSpeech, trains three times faster than LSTM, halves model size, and dramatically speeds up real‑time decoding.
