kul-speech-lab/contextual-reasoning-speechllm

Contextual Reasoning for Speech-LLM

This dataset accompanies the paper

Towards Deep Contextual Reasoning from Broad Descriptions for ASR with Speech-LLM via Metadata-Driven Reasoning Chains
Jakob Poncelet and Hugo Van hamme. Interspeech, 2026.

It bundles training and evaluation data for reasoning-augmented speech-LLM ASR models. Each row pairs a speech segment with the LLM-generated reasoning chain used to correct ASR errors using broad textual context (title, description, tags) of the underlying recording.

The raw audio is not redistributed. Every row carries the information needed to recover the audio from the publicly available source dataset. See Audio retrieval.

---

Configurations (Subsets)

The dataset is organised as a single repository with 6 configurations selectable from the Subset dropdown:

config	rows	source mix	purpose
train_l	299 906	gigaspeech 71% / slidespeech 25% / slideavsr 4%	Full training set
train_m	129 484	gigaspeech 74% / slidespeech 22% / slideavsr 4%	LLM-filtered training subset
train_s	27 125	gigaspeech 100%	Gigaspeech rows containing rare named entities
test_m3av	3 914	m3av 100%	M3AV evaluation set
test_slideavsr	2 099	slideavsr 100%	SlideAVSR evaluation set (testA)
test_slidespeech	3 188	slidespeech 100%	SlideSpeech evaluation set (test)

Each configuration is a single denormalized table: one row per (segment, ASR-model hypothesis) pair. The same segment can appear several times because it is paired with hypotheses from whisper_tiny, whisper_base, whisper_small, whisper_large-v3, and a None bucket (reference-only / artificial-error examples).

---

Schema

Every config has the same columns (with one optional structured_reasoning column present in the training configs):

Identifiers and source

column	type	description
id	string	Primary key. Formatted as <seg_id>__<source_model> (or just <seg_id> when source_model is null).
seg_id	string	Segment identifier. Encodes the source's own utterance ID.
file_id	string	Recording / audio file identifier in the source dataset.
source	string	One of gigaspeech, m3av, slidespeech, slideavsr.
source_model	string or null	Whisper variant whose hypothesis is used here (whisper_tiny, whisper_base, whisper_small, whisper_large-v3), or null for reference-only / artificial-error examples.

Transcripts

column	type	description
text	string	Reference transcript for the segment (normalised by the upstream toolchain).
hypothesis	string	ASR hypothesis from source_model.

Reasoning

column	type	description
reasoning	string	Free-text LLM reasoning derived from the structured analysis.
structured_reasoning	string (JSON)	LLM's structured analysis of the hypothesis vs reference. Only present in train_l and train_m.
corrections	string (JSON list)	Filtered list of token-level corrections proposed by the reasoning chain.
corrections_all	string (JSON list)	Unfiltered list of all proposed corrections.
num_corrections	int64 or null	Count of kept corrections.
dropped_audio_only_corrections	int64 or null	Count of corrections that were dropped because they were judged "audio-only" (not justified by the textual context).

Segment timing / provenance

column	type	description
duration	float	Length of the segment audio used during training, in seconds.
start	float	Start time of the segment in the original source recording, in seconds.
end	float	End time of the segment in the original source recording, in seconds.
original_segment_ids	list[string]	For gigaspeech rows: the list of original (pre-merge) gigaspeech segment IDs that were concatenated to produce this segment. Empty list for non-gigaspeech rows.
source_segments_path	string	Provenance label identifying the upstream segmentation split (e.g. speechcolab/gigaspeech (subset=l), slidespeech/long/L95/segments).

Gigaspeech caveat. For source == "gigaspeech" rows, the audio used at training time is the concatenation of the original gigaspeech segments listed in original_segment_ids, with inter-segment silences removed. The start / end columns give the span min(begin) / max(end) across constituent segments and therefore include the removed pauses; do not extract [start, end] as a single slice — extract each original_segment_ids segment separately and concatenate them. See Audio retrieval — gigaspeech.

Context corpus

column	type	description
title	string	Title (e.g. YouTube title for YouTube-sourced rows, paper/slide title for m3av).
description	string	Cleaned LLM-generated description of the recording's content.
tags	list[string]	Topical / entity tags.

The title/description/tags fields are the "context corpus" that the LLM reasoning chain conditions on for context-aware ASR error correction. These fields repeat across rows of the same recording.

---

Usage

from datasets import load_dataset

REPO = "kul-speech-lab/contextual-reasoning-speechllm"

ds = load_dataset(REPO, "train_l")           # full training set
# DatasetDict({train: Dataset({features: [id, seg_id, file_id, ...], num_rows: 299906})})

# Each row is one (segment, hypothesis) example with everything joined inline
ex = ds["train"][0]
print(ex["id"], ex["source"], ex["source_model"])
print("REF :", ex["text"])
print("HYP :", ex["hypothesis"])
print("REAS:", ex["reasoning"])
print("CTX :", ex["title"], "|", ex["tags"][:3])
print("audio span: [", ex["start"], ",", ex["end"], "] in source", ex["file_id"])

To filter to a single ASR source:

giga_only = ds["train"].filter(lambda x: x["source"] == "gigaspeech")

To deduplicate to one row per segment (regardless of which ASR hypothesis):

seen = set()
unique_segments = ds["train"].filter(lambda x: x["seg_id"] not in seen and not seen.add(x["seg_id"]))

---

Audio retrieval

The dataset does not ship audio. Each row's (file_id, start, end) (and, for gigaspeech, original_segment_ids) gives you everything needed to extract the matching audio from the public source.

Source identification

The source column tells you which upstream dataset to query.

source	upstream dataset	license	gated?
gigaspeech	speechcolab/gigaspeech (subset l)	SpeechColab user-agreement (non-commercial research/education)	yes — sign their Google Form
m3av	CHHHH/M3AV_v1.0	CC BY-NC 4.0	no
slidespeech	openslr 144 (SlideSpeech) — and the wavid2channel mapping from https://slidespeech.github.io/	Apache 2.0	no
slideavsr	nlp-waseda/SlideAVSR (videos downloaded from YouTube via their dataset/reproduce.sh)	CC BY-NC-SA 4.0	no

Retrieval procedure

For m3av, slidespeech, slideavsr — every row's [start, end] is a contiguous slice of the source recording, so a single ffmpeg call suffices:

ffmpeg -y -i {file_id}.wav -ss {start} -to {end} -ac 1 -ar 16000 {seg_id}.wav

The file_id is the recording id in the upstream dataset (YouTube video id for slideavsr/m3av; named id like child_0000 for slidespeech). Refer to the upstream dataset's documentation for how to locate the wav file.

For gigaspeech — the segment used in training is a concatenation of multiple original gigaspeech segments, with inter-segment silences removed by the merge step. A single [start, end] slice will not match. Use the original_segment_ids list to look up each constituent segment in speechcolab/gigaspeech (subset l), extract its [begin_time, end_time] from the source audio, and concatenate the slices in order:

from datasets import load_dataset
import numpy as np

giga = load_dataset("speechcolab/gigaspeech", "l", split="train")  # gated; also check validation+test
by_id = {r["segment_id"]: r for r in giga}

def recover_gigaspeech_audio(row):
    parts = []
    sr = None
    for oid in row["original_segment_ids"]:
        src = by_id[oid]
        arr = src["audio"]["array"]
        sr  = src["audio"]["sampling_rate"]
        parts.append(arr)
    return np.concatenate(parts), sr

---

License

This release packages four upstream datasets, each with its own license. You must satisfy all four licenses for the rows you use.

source	license	gated?
Gigaspeech	SpeechColab user-agreement (non-commercial research/education)	yes
M3AV	CC BY-NC 4.0	no
SlideSpeech	Apache 2.0 (openslr release)	no
SlideAVSR	CC BY-NC-SA 4.0	no

The non-audio fields in this repository — reasoning chains, structured reasoning, corrections, file descriptions, tags — are LLM-generated derivatives of the source transcripts and metadata. They are released for non-commercial research use, consistent with the strictest upstream license (CC BY-NC-SA 4.0).

---

Citation

@inproceedings{poncelet2026contextual,
  title     = {Towards Deep Contextual Reasoning from Broad Descriptions for ASR with Speech-LLM via Metadata-Driven Reasoning Chains},
  author    = {Poncelet, Jakob and Van hamme, Hugo},
  booktitle = {Proc. Interspeech},
  year      = {2026}
}