htx-language-ai-alpha/htd-multilingual-spoken-qa

HTD Multilingual Spoken QA Dataset

A spoken-input, text-output multi-turn QA dataset covering Singapore Home Team domains across English, Mandarin Chinese, Bahasa Melayu, and Tamil, with natural Singaporean colloquial register. This is the audio companion to htd-multilingual-text-qa: every user turn is replaced by a voice-cloned TTS rendering of the corresponding user text, while assistant turns remain text in the target language.

Overview

Each row is a complete user-assistant conversation about a Home Team topic. Odd-indexed turns (turn_1, turn_3, …) are user audio — WAV-encoded voice-cloned TTS waveforms generated from the upstream conversation dataset's user utterances. Even-indexed turns (turn_2, turn_4, …) are assistant text in the target language. The dataset is intended for training/evaluating small speech-language models on multilingual spoken QA grounded in Singapore Home Team subject matter.

Audio is stored as a native HuggingFace datasets.Audio feature (FLAC-encoded, 16 kHz mono), so the odd turn_* columns decode automatically on access via load_dataset (no cast_column needed) and play inline in the Hub dataset viewer. The FLAC encoding is lossless — bit-exact to the original 16-bit PCM — and roughly halves the on-disk size versus the raw WAV.

Source domains (the collection column) cover four Home Team agencies:

Collection	Agency	Rows
MHA	Ministry of Home Affairs	174,359
SCDF	Singapore Civil Defence Force	169,782
SPF	Singapore Police Force	101,378
HTX	Home Team Science & Technology Agency	59,334

Statistics

Total: 504,853 rows, distributed roughly evenly across the four languages (each conversation is assigned exactly one language via round-robin during generation in the upstream text dataset):

Language	Rows
English (en)	126,300
Bahasa Melayu (ms)	126,299
Mandarin Chinese (zh)	126,299
Tamil (ta)	125,955
Total	504,853

Across the whole dataset:

• User-audio turns: 1,724,306
• Assistant-text turns: 1,723,693

(The ~600-turn gap between user-audio and assistant-text totals comes from a ~0.24% tail of upstream rows that carry one extra "stranded" user turn beyond 2 * actual_num_turns — see the upstream README. The 344-row Tamil shortfall vs. the other languages mirrors the upstream conversation dataset.)

Audio duration

User-turn audio is 16 kHz mono, stored losslessly as FLAC (bit-exact to the original 16-bit PCM). Across all 1,724,306 user-audio cells (post-repair):

Stat	Value
Minimum duration	1.25 s
Maximum duration	89.99 s
Mean duration	~10.8 s
Total audio	~5,190 hours

Every expected-populated user-audio cell decodes to a non-empty waveform; see steps 7–8 of "How the dataset was developed" for the post-hoc repairs that ensured this and that capped the long tail.

Conversation length distribution (QA pairs per row; actual_num_turns):

QA pairs	Rows
2	130,863
3	136,923
4	137,904
5	96,105
6	2,963
7	89
8	6

Note on actual_num_turns: despite its name, this column counts QA pairs, not raw turn count. A row with actual_num_turns = 3 has 6 populated turn_* columns (3 user audio + 3 assistant text), not 3. To iterate the full conversation, use range(1, 2 * actual_num_turns + 1). A small tail (~0.24%) carries one extra stranded user-audio turn beyond 2 * actual_num_turns.

Schema

Column	Type	Description
unique_id	str	Conversation id from the upstream conversation dataset; not row-unique
collection	str	Source agency: MHA / SCDF / SPF / HTX
problem_statement	str	Topical seed prompt the conversation is grounded in
persona	str	Singaporean speaker persona used for the user turns
language	str	Language of the entire conversation: en / ms / ta / zh
num_turns	int	Target number of QA pairs requested at generation time (2–5)
actual_num_turns	int	Number of QA pairs actually produced (2–8). The full conversation has 2 * actual_num_turns populated turns; see Statistics above.
turn_1, turn_3, … turn_15	HF Audio feature	User-turn audio. 16 kHz mono, lossless FLAC (stored as {bytes, path}; auto-decodes to a waveform array via datasets). Unused turns are null.
turn_2, turn_4, … turn_16	str	Assistant-turn text in the target language. Unused turns are null.
checker_result	str	Quality-check result from the upstream generation pipeline (typically Y)

How the dataset was developed

1. Upstream conversations. All rows come from htd-multilingual-text-qa (504,853 rows). User and assistant text were generated as multi-turn QA exchanges grounded in Singapore Home Team source material (MHA / SCDF / SPF / HTX), then translated whole-conversation into one of en / zh / ms / ta round-robin, with user turns in casual code-switched Singaporean register and assistant turns in the target language. See the upstream README for the text pipeline (generation, code-switched translation, script guardrails, Tamil repair pass, final regex cleanup).

2. Voice-cloned TTS for user turns. Odd-indexed (user) turns were synthesized with the local OmniVoice model (k2-fsa/OmniVoice) running in-process via text_audio_augmentation/voice_clone_tts.py. For each row a reference voice is picked at random from the per-language pool:

   • en — en_male / en_female
   • zh — zh_male / zh_female
   • ms — ms_male_1 / ms_male_2
   • ta — ta_male / ta_female

   Reference clips and their transcripts live in text_audio_augmentation/reference_audio/. The references are themselves Singapore-style code-switched utterances, so the cloned voices retain the casual code-switched register of the source text (Singlish particles, English nouns/acronyms in Latin script, etc.).

3. Text normalization before TTS. User text passes through light normalization to help the TTS pronounce it: Arabic numerals are spelled out (English num2words), and all-caps acronyms (SCDF, HTX, NSF, …) are space-separated so they are read letter-by-letter rather than as fake words.

4. Audio encoding. OmniVoice output is 16 kHz mono 16-bit PCM. For distribution it is re-encoded losslessly to FLAC and stored as a native HuggingFace datasets.Audio feature ({bytes, path}), so the columns auto-decode via load_dataset and play inline in the Hub viewer. The FLAC round-trip is bit-exact to the source PCM (verified sample-for-sample), and shrinks the dataset by 49% (527 GB WAV → ~258 GB FLAC) with zero quality loss and 16 kHz preserved.

5. Assistant turns unchanged. Even-indexed assistant turns are passed through as text from the upstream conversation dataset.

6. Sharding. The full dataset is split into 113 parquet shards of ~5 GB each for easier handling and HF-style loading. Each shard holds ~4,500 rows.

7. Post-hoc repair of empty WAVs. A full audit found exactly one user-audio cell across the dataset whose WAV decoded to zero frames (train-00096-of-00113.parquet row 2938 turn_7, unique_id = scdf_987, language = en). The cause was upstream: the source text in htd-multilingual-text-qa at that cell was literally '...', so OmniVoice had nothing to synthesize. The fix touched both datasets — the user turn was regenerated with Azure OpenAI from the surrounding conversation (turns 1–6 plus the existing turn_8 assistant reply that constrains what the question must have been), the text-qa cell was rewritten in place, and the new text was re-synthesized through the same OmniVoice pipeline (en_female reference) to produce a real 3.68 s WAV.

8. Post-hoc repair of the long tail. A second audit found 392 user-audio cells whose duration exceeded 90 s — 370 of them in turn_1, mostly Tamil / Malay / Mandarin. Inspecting the upstream text showed the same root cause across all of them: the LLM that generated user turns occasionally produced a 1,500–2,600-character assistant-style explanation ("நான் சுலபமா explain பண்ணுறேன்…", "Boleh ah, saya explain sikit…", "关于…的规则…") in a user slot. OmniVoice faithfully turned those into 90–173 s utterances. The repair followed the same two-sided pattern as step 7: each affected user turn was regenerated with Azure OpenAI gpt-4.1-mini given the row's persona, problem statement, and full surrounding conversation (the assistant reply that the new question must lead into is the strongest constraint), the text-qa cells were rewritten in a single atomic pass, and the new texts were re-synthesized through OmniVoice with the same per-language reference-voice picker. A confirming re-scan reported 0 user-audio cells longer than 90 s and a new maximum of 89.99 s. 8,262 cells between 60 s and 90 s remain (mostly long but real turn_1 questions); they were not modified.

Examples

Each example below is the audio counterpart of the matching example in the upstream text-qa README — same SCDF row, same actual_num_turns = 4. User turns are audio in this dataset; the user text shown is the source string that the voice-cloned TTS rendered for that cell.

English (en)

Persona: A Singaporean youth advocate with a diploma in social work, pushing for mental health resources in schools.

Problem statement: What initiatives have been implemented to improve the efficiency, training, and welfare of National Servicemen (NS) in Singapore?

turn_1 (user audio of): "Hey, I'm curious, how does the Singapore Police Force ensure NS officers are trained to handle stressful operations like public disorder situations?"

turn_2 (assistant): The Singapore Police Force has introduced specialized training for NS officers to prepare them for operational readiness in handling public disorder. Starting from 2018, two Public Order Troops within the Special Operations Command (SOC) are formed annually, equipping NS Full-time (NSF) officers with the same capabilities as regular Police Tactical Troops…

turn_3 (user audio of): "Wow, that sounds intense! But how do they make sure the NS officers actually perform well under all that pressure?"

turn_4 (assistant): To ensure NS officers perform well under pressure, the training incorporates scenario-based exercises designed to mimic high-stress environments…

Bahasa Melayu (ms)

Persona: A Singaporean management professional with a bachelor's degree, focused on optimizing resources and implementing patient-centered care initiatives.

Problem statement: What are the minimum concrete dimensions required to achieve various levels of fire resistance for reinforced concrete structures?

turn_1 (user audio of): "Saya curious sikit, jenis aggregate dalam concrete tu main peranan apa bila nak tentukan fire resistance requirement untuk structural columns?"

turn_2 (assistant): Jenis agregat yang digunakan dalam konkrit memberi kesan yang ketara kepada keperluan ketahanan api untuk tiang struktur. Contohnya, konkrit agregat silika yang kurang stabil secara termal biasanya memerlukan dimensi yang lebih besar untuk mencapai ketahanan api yang sama berbanding konkrit agregat batu kapur atau agregat ringan…

turn_3 (user audio of): "Oh, itu menarik. Kalau saya nak guna lapisan pelindung tambahan macam gypsum plaster, itu boleh bagi kesan yang bermakna ke pada saiz kolum?"

turn_4 (assistant): Ya, menambah lapisan pelindung seperti plaster gypsum boleh memberikan perbezaan yang ketara. Untuk tiang konkrit agregat silika dengan semua muka terdedah, menggunakan plaster gypsum setebal 15 mm pada pengukuhan jaring ringan membolehkan ketebalan yang diperlukan untuk ketahanan api selama 4 jam dikurangkan daripada 450 mm kepada 400 mm…

Tamil (ta)

Persona: A Singaporean undergraduate studying computer science, passionate about developing AI-driven solutions to enhance urban living in Singapore.

Problem statement: What are the roles and responsibilities of key personnel within the SCDF's Operation Lionheart contingent during overseas missions?

turn_1 (user audio of): "ஒரு mission க்கு deploy பண்ணும்போது, SCDF எப்படி எல்லா equipment களும் resources களும் நேரத்துக்கு தயார் பண்ணிக்கோறாங்க?"

turn_2 (assistant): SCDF-வின் Operation Lionheart குழுவில் Logistics Officer-ன் பொறுப்பு, ஒரு பணிக்கான அனைத்து உபகரணங்களும் வளங்களும் தயார் நிலையில் இருப்பதை உறுதி செய்வதாகும். அனுப்புவதற்கு முன், Logistics Officer அவசியமான உபகரணங்களை பராமரித்து, தேவையான அனைத்து ஆவணங்களும் புதுப்பிக்கப்பட்டுள்ளதா என கவனிக்கிறார்…

turn_3 (user audio of): "அவர்கள் disaster zone-க்கு வந்து சேர்ந்தப்போது, எப்படி base-ஐ அமைத்து team-க்கு எல்லாமே ஒழுங்கு பண்ணுவாங்க?"

turn_4 (assistant): விபத்து மண்டலத்திற்கு வந்தவுடன், Logistics Officer Base of Operations அமைக்க ஏற்ற இடத்தை கண்டறிதலில் பொறுப்பேற்கிறார். அவர்கள் பாதுகாப்பு நடவடிக்கைகளை நிறுவி, base உள்ள பொதுவான வாழ்வுக்கான விதிகளை நியமிக்கிறார்…

Mandarin Chinese (zh)

Persona: A Singaporean real estate agent with a diploma in real estate management, specializing in promoting eco-friendly housing options.

Problem statement: What resources are available for Fire Safety Managers (FSMs) to improve fire safety knowledge and practices?

turn_1 (user audio of): "嗨，我想问一下，有没有什么presentation或者resources，是讲消防安全经理在检查时常遇到的问题的？can share一下吗？"

turn_2 (assistant): 有一些资源可以解决这些问题。例如，有一份由CPT Kang Kim Long主讲的名为"Common Fire Safety Enforcement Findings and Update on Handling of Fire Safety Feedback"的演示文稿。该文件将提供关于常见执法发现以及如何处理消防安全反馈的见解。

turn_3 (user audio of): "哦，这听起来蛮有用的。你知道有没有什么关于假警报的情况或者例子，还有怎么处理它们的吗？"

turn_4 (assistant): 是的，有一个专门针对该主题的资源。由David Goh先生主讲的题为"FSM Perspective on False Alarms and Practical Approach to Fire Alarm Activation"的演讲，提供了关于应对误报以及管理火警警报激活的实用策略的见解。

Loading

from datasets import load_dataset

ds = load_dataset(
    "parquet",
    data_files="train-*-of-00113.parquet",
    split="train",
)
# Odd turn_* columns are already an Audio feature — no cast_column needed.
# They decode to a waveform dict on access:
audio = ds[0]["turn_1"]          # {"array": np.ndarray, "sampling_rate": 16000, "path": None}

For very large iteration without populating the HF arrow cache:

ds = load_dataset(
    "parquet",
    data_files="train-*-of-00113.parquet",
    split="train",
    streaming=True,
)

Or directly with PyArrow (no HF cache, no audio decoding):

import pyarrow.parquet as pq
table = pq.ParquetDataset(".").read()

Files

• train-00000-of-00113.parquet … train-00112-of-00113.parquet — 113 parquet shards, 2.6 GB each (258 GB total), totaling 504,853 rows. Written with small row groups (~400 rows) so the Hub viewer renders without TooBigContentError.
• README.md — this file