Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Alignment-Free Training for Transducer-based Multi-Talker ASR

Takafumi Moriya, Shota Horiguchi, Marc Delcroix, Ryo Masumura, Takanori Ashihara, Hiroshi Sato, Kohei Matsuura, Masato Mimura

TL;DR

A novel alignment-free training scheme for the MT-RNNT (MT-RNNT-AFT) that adopts the standard RNNT architecture that achieves performance comparable to that of the state-of-the-art alternatives, while greatly simplifying the training process.

Abstract

Extending the RNN Transducer (RNNT) to recognize multi-talker speech is essential for wider automatic speech recognition (ASR) applications. Multi-talker RNNT (MT-RNNT) aims to achieve recognition without relying on costly front-end source separation. MT-RNNT is conventionally implemented using architectures with multiple encoders or decoders, or by serializing all speakers' transcriptions into a single output stream. The first approach is computationally expensive, particularly due to the need for multiple encoder processing. In contrast, the second approach involves a complex label generation process, requiring accurate timestamps of all words spoken by all speakers in the mixture, obtained from an external ASR system. In this paper, we propose a novel alignment-free training scheme for the MT-RNNT (MT-RNNT-AFT) that adopts the standard RNNT architecture. The target labels are created by appending a prompt token corresponding to each speaker at the beginning of the transcription, reflecting the order of each speaker's appearance in the mixtures. Thus, MT-RNNT-AFT can be trained without relying on accurate alignments, and it can recognize all speakers' speech with just one round of encoder processing. Experiments show that MT-RNNT-AFT achieves performance comparable to that of the state-of-the-art alternatives, while greatly simplifying the training process.

Alignment-Free Training for Transducer-based Multi-Talker ASR

TL;DR

A novel alignment-free training scheme for the MT-RNNT (MT-RNNT-AFT) that adopts the standard RNNT architecture that achieves performance comparable to that of the state-of-the-art alternatives, while greatly simplifying the training process.

Abstract

Extending the RNN Transducer (RNNT) to recognize multi-talker speech is essential for wider automatic speech recognition (ASR) applications. Multi-talker RNNT (MT-RNNT) aims to achieve recognition without relying on costly front-end source separation. MT-RNNT is conventionally implemented using architectures with multiple encoders or decoders, or by serializing all speakers' transcriptions into a single output stream. The first approach is computationally expensive, particularly due to the need for multiple encoder processing. In contrast, the second approach involves a complex label generation process, requiring accurate timestamps of all words spoken by all speakers in the mixture, obtained from an external ASR system. In this paper, we propose a novel alignment-free training scheme for the MT-RNNT (MT-RNNT-AFT) that adopts the standard RNNT architecture. The target labels are created by appending a prompt token corresponding to each speaker at the beginning of the transcription, reflecting the order of each speaker's appearance in the mixtures. Thus, MT-RNNT-AFT can be trained without relying on accurate alignments, and it can recognize all speakers' speech with just one round of encoder processing. Experiments show that MT-RNNT-AFT achieves performance comparable to that of the state-of-the-art alternatives, while greatly simplifying the training process.
Paper Structure (15 sections, 2 equations, 2 figures, 2 tables)

This paper contains 15 sections, 2 equations, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Baseline systems
  4. Single-talker RNNT (Standard RNNT)
  5. MT-RNNT with tSOT (MT-RNNT-tSOT)
  6. Proposed methods
  7. Alignment-free training for MT-RNNT (MT-RNNT-AFT)
  8. Self-knowledge distillation for MT-RNNT-AFT
  9. Experimental evaluations
  10. Data
  11. System configuration
  12. Offline results
  13. Streaming results
  14. Effect of beam size in inference of each MT-RNNT
  15. Conclusion

Figures (2)

  • Figure 1: Training procedure of MT-RNNT-tSOT Kanda2022mtasr. Token-level serialized transcription $Y^{\text{tSOT}}$ is generated from both speakers' transcriptions and their forced alignments obtained from an external ASR system. Both $Y^{\text{tSOT}}$ and its mixture, $\bm{X}^{\text{mixture}}$, are generated on-the-fly with a random delay.
  • Figure 2: Training procedure of MT-RNNT-AFT. MT-RNNT-AFT decodes all speakers' speech in a first-in-first-out manner. Prompt tokens <spk$m$>, which correspond to the sequential order of each speaker’s appearance in mixture $\bm{X}^{\text{mixture}}$, are appended to the beginning of each transcript $Y^{\text{spk}m}$.