sg-podcast-sqa — Spoken-Question-Answering from Singaporean Podcasts

Spoken-Question-Answering (SQA) data for a speech LLM (MERaLiON), generated from Singaporean podcasts with gpt-4o-audio. Each row is an (audio segment, question, answer) triple: the model hears an audio chunk and must answer the question in text. The audio is sourced from malaysia-ai/singaporean-podcast-youtube, and the question taxonomy follows mispeech/MECAT-QA.

The dataset is split by podcast (grouped on source_file), so no podcast appears in both train and test — a model trained on the train split has never heard the audio it is evaluated on, preventing audio leakage into evaluation. The held-out podcast set is identical to the previous version of the dataset, so evaluations stay comparable across versions even though the questions and answers are regenerated.

Answers are reasoning-rich. Every answer states the answer, then cites the specific audio evidence and the reasoning that connects the evidence to the answer (e.g. "the first speaker says X, which shows Y"). This is a deliberate change from an earlier terse version: a model SFT'd on short answers under-performed the baseline on answer detail, so the ground-truth targets now teach detailed, well-justified responses (~4–7 sentences; ~71 words on average) while keeping the concrete specifics (names, numbers, quotes, who-said-what).

Sources

• Audio — malaysia-ai/singaporean-podcast-youtube: Singaporean podcast audio from YouTube, collected by Malaysia AI.
• Question taxonomy — mispeech/MECAT-QA: the audio-QA category / difficulty scheme the question types are based on.

Splits

Split	Rows	Podcasts (source_file)	Shards	Share	Mean answer
train	150,654	3,266	66 (train-00000-of-00066.parquet …)	94.96%	~71 words
test	8,000	169	4 (test-00000-of-00004.parquet …)	5.04%	~72 words
Total	158,654	3,435	70	100%	—

Train and test share zero podcasts.

Layout

Both splits reference a single shared audio/ folder at the dataset root:

sg-podcast-sqa/          <- point your loader / audio resolver here
├── audio/                  <- all FLAC chunks (6,072 files, shared by train AND test)
│   └── <podcast>__chunkNNN.flac
├── train/
│   └── train-*.parquet     (66 shards)
├── test/
│   └── test-*.parquet      (4 shards)
└── README.md               <- this dataset card

audio_path is stored relative to the dataset root (sg-podcast-sqa/), e.g. audio/My_Podcast_id__chunk000.flac — not relative to a split folder. When you load a parquet directly, join audio_path against the dataset root, not against train/ or test/.

Audio format. The audio is lossless FLAC, 16 kHz mono (~85 GB total).

Each audio chunk is up to ~20 min; a podcast is split into n_chunks chunks and each chunk yields many QA pairs.

Schema

column	type	description
audio_path	string	FLAC path relative to the dataset root, e.g. audio/<podcast>__chunk000.flac
sampling_rate	int32	audio sampling rate (Hz) — always 16000
instruction	string	the question posed about the audio (English)
answer	string	the ground-truth answer (English) — states the answer plus its audio evidence and reasoning
question_type	string	one of the 9 task types below
level	string	difficulty: basic, intermediate, advanced, complex
content_language	string	language of the speech the question is about: en, zh, ms, ta, mixed
source_file	string	the source podcast (the split key — no podcast crosses train/test)
chunk_idx / n_chunks	int32	this chunk's index / total chunks for the podcast
chunk_start_sec / chunk_end_sec	float64	this chunk's span within the full podcast
generation_status	string	always success (non-success rows were filtered out)
row_idx	int64	0-based row index, re-sequenced per shard

The 9 question types (tasks)

These categories are based on the mispeech/MECAT-QA taxonomy, extended here with speech_content, summarization, and colloquial.

Each answer surfaces the concrete specifics (named people/places/orgs, numbers, prices, dates, direct claims, who-said-what) and the justification — the audio evidence and the reasoning linking it to the answer. Real examples from the dataset:

speech_content — the substance of what is said (topics, facts, numbers, opinions, who said what). The dominant type (~46–48%).

Q: What is the title of the podcast episode and how did the first speaker introduce it? A: The title of the podcast episode is not explicitly stated, but the first speaker introduces it as the first episode ever of the 'Takeaway Table Podcast.' He says, 'Hey everyone, welcome to the first episode ever of the Takeaway Table Podcast,' which directly identifies the name and the fact that this is their inaugural episode.

summarization — a thorough summary at some scope (whole segment, one portion, one speaker's points, or the takeaways). The longest answers, enumerating specific points.

Q: Summarize the speakers' explanation of why they are recording this podcast episode. A: The speakers explain that they are recording this podcast episode to answer a frequently asked question about how they got started on social media. They mention that they have been making online digital content for seven years and often get asked how they began. They want to have a recorded answer to refer people to … They introduce themselves, joke about sounding alike, and discuss their personal lives, including one speaker having a child in 2019. …

inference_judgment — reasoning beyond the literal content (mood, relationships, implications), still anchored to what was said and why.

Q: How do the speakers describe their childhood interest in entertainment, and what does this imply about their relationship? A: The speakers describe their childhood interest in entertainment by recalling that they were involved in church plays, singing, and Christmas productions. They mention that they always wanted to direct the plays themselves … This implies a close relationship and shared passion from a young age, as they collaborated on creative projects and had similar interests, which likely strengthened their bond.

application_context — the real-world meaning/purpose of what is discussed, with the evidence for it.

Q: Why do the speakers mention 'Inside Scoop' and what does this reveal about their approach to the podcast? A: They mention 'Inside Scoop' jokingly as a non-sponsor, suggesting that if the ice cream brand wants to sponsor them, they would welcome it. This reveals that they are lighthearted and open to sponsorships, and it shows their playful approach to the podcast, as they weave humor into their discussion.

colloquial — only when speakers use Singlish or Malay/Mandarin/Tamil terms; quotes the actual term, explains it, and cites where it was used.

Q: How do the speakers use the term 'cooking pot' metaphorically, and what does it mean in this context? A: They say the podcast has been in the 'cooking pot' for a long time, which metaphorically means it has been an idea they have been simmering and developing. The first speaker says, 'It's been in the cooking pot for a long time, like too long,' indicating that they have been planning or wanting to do this podcast for a while but only now are actually doing it.

direct_perception — identify sounds/voices/what is audibly happening, with the evidence.

Q: How many speakers are present in the audio and what evidence shows this? A: There are two speakers present in the audio. The evidence is that the first speaker greets the audience and introduces himself as Ming, and then the second speaker also introduces himself as Ming. They acknowledge that they both sound similar and joke about it, indicating two distinct voices taking turns to speak.

sound_characteristics — properties of the sound (tone, pace, pitch, music vs speech, number of voices), with what in the audio shows it.

Q: What is the tone and pace of the speakers' conversation, and what in the audio shows this? A: The tone is casual and humorous, and the pace is moderate. They joke about needing a name for the podcast and about sounding alike, which shows a relaxed tone. The pace is steady, with each speaker taking turns to speak without rushing, allowing the audience to follow along easily.

quality_assessment — recording quality (clarity, noise, distortion), with the audible basis for the judgement.

Q: Assess the recording quality of the podcast based on what is heard. A: The recording quality is high, with clear and crisp voices and no noticeable background noise or distortion. The speakers' voices are balanced in volume, and there is no echo, suggesting the use of good microphones and a controlled recording environment. This clarity allows listeners to easily distinguish between the two speakers even though they sound similar.

environment_reasoning — infer the setting / acoustic environment from the evidence.

Q: What evidence in the audio suggests that the speakers are recording in a controlled indoor environment? A: The audio is clear, with no background noise like traffic or wind, and the voices sound close to the microphone, indicating a controlled indoor environment. The speakers also mention that they will upload the audio online and talk about linking it to YouTube or Facebook, suggesting they are in a studio-like setting where recording equipment is set up.

Distribution by question type

question_type	train rows	train %	test rows	test %
speech_content	68,994	45.8%	3,808	47.6%
inference_judgment	22,581	15.0%	1,196	14.9%
summarization	14,931	9.9%	775	9.7%
application_context	10,964	7.3%	557	7.0%
colloquial	10,086	6.7%	510	6.4%
sound_characteristics	7,247	4.8%	387	4.8%
environment_reasoning	5,955	4.0%	289	3.6%
quality_assessment	4,966	3.3%	258	3.2%
direct_perception	4,930	3.3%	220	2.8%

Language (content_language): en ≫ ms > mixed, with zh/ta ≈ 0 — the source corpus is English + Malay + Singlish, so Mandarin/Tamil are nearly absent. Level: advanced ≈ intermediate > basic > complex. Both mixes are consistent across the two splits.

Loading with 🤗 datasets

Audio is stored by reference (audio_path is a string pointing at a FLAC file under audio/, not an embedded Audio feature), so load the parquet and resolve audio_path against the dataset root yourself:

from datasets import load_dataset

ROOT = ".../sg-podcast-sqa"   # dataset root (has audio/, train/, test/)

ds = load_dataset("parquet", data_files={
    "train": f"{ROOT}/train/train-*.parquet",
    "test":  f"{ROOT}/test/test-*.parquet",
})

import os, soundfile as sf
row = ds["test"][0]
audio_path = os.path.join(ROOT, row["audio_path"])   # resolve relative to ROOT (a .flac file)
audio, sr = sf.read(audio_path)                       # soundfile decodes FLAC natively
print(row["instruction"], "->", row["answer"])

To decode audio automatically via datasets, add the column and cast it:

from datasets import Audio

ds = ds.map(lambda r: {"audio": os.path.join(ROOT, r["audio_path"])})
ds = ds.cast_column("audio", Audio(sampling_rate=16000))
ds["test"][0]["audio"]   # {'array': ..., 'sampling_rate': 16000, 'path': ...}

Note on relative paths. load_dataset does not know your ROOT, so audio_path stays a plain relative string — you must os.path.join(ROOT, audio_path). If your pipeline resolves audio relative to a split's own folder instead of the dataset root, rewrite the paths to ../audio/… first.