StreamAlign: Streaming Text-Aligned Speech Tokenization

Speech Reconstruction

Sample	Ground Truth	Streaming, text-aligned	Non-streaming		Streaming			Text-informed
Sample	Ground Truth	StreamAlign (Ours)	SpeechTokenizer	WavTokenizer	Mimi	MagiCodec	Qwen3-TTS	TASTE	TaDiCodec
Sample 1 Three seconds before the arrival of J. B. Hobson's letter, I no more dreamed of chasing the unicorn than of trying for the Northwest Passage.
Sample 2 What do you mean? inquired Louis.
Sample 3 Don't I though, I'm so sorry to hear it, how did her son turn out?
Sample 4 I have a letter here, Mister Soames, which I wrote to you early this morning in the middle of a restless night.
Sample 5 Oh my Lord, cried Miss Woodley with a most forcible accent, you are the last person on earth she would pardon me for entrusting.

Sample	Prompt	StreamAlign-SLM (Ours) Llama-3.2 1B	ASR + LLM + TTS Llama-3.2 1B	Spoken Language Model Baselines
Sample	Prompt	StreamAlign-SLM (Ours) Llama-3.2 1B	ASR + LLM + TTS Llama-3.2 1B	TWIST OPT 1.3B	TWIST Llama 7B	SpiRit-LM Llama-2 7B	SpiRit-LM (expr.) Llama-2 7B	TASLM Llama-3.2 1B
Sample 1 “there is no enchantment nor charm there”
Sample 2 “Did you ever have such a lordly guest before?”
Sample 3 “you stick to that said beer radiant with delight”
Sample 4 “I do not know.”
Sample 5 “She gathered up her reins.”