ASR technology excels in accurately transcribing prepared read speech, but it still encounters challenges when dealing with spontaneous conversations. This is partly because the latter is an unprepared,casual language register with lots of disfluencies and speech reductions. In order to improve ourknowledge about speech variations, we designed an oral corpus of 36 hours of Mandarin Chinese.Spontaneous conversations are automatically transcribedand selected excerpts of these are then readby their authors. We employed WHISPER for automatic speech transcription and the Montreal ForcedAligner for forced alignment, which represents an effective procedure from speech to text, then tomulti-tier annotation within phones, characters/syllables, and words, particularly suitable for largescale speech corpus construction. This enabled us to collect different speaking styles with partlyoverlapped content produced by the same speaker. Such parallel corpus are particularly helpful toinvestigate temporal speech reduction phenomena in spontaneous speech.