VoxCanvas-41M

A Large-Scale Bilingual Dataset for Instruction-Based Speech Editing and Its Learning Dynamics

VoxCanvas-41M is a large-scale bilingual (Chinese and English) dataset for instruction-based speech editing. It contains about 41 million editing pairs built from real speech, in two categories:

• Content editing: single-span editing, multi-span editing (two to six edits in one instruction), and location-reference editing, where the edit is located by occurrence, sentence boundary or timestamp.
• Paralinguistic editing: insertion and deletion of 39 classes of nonverbal vocalizations, such as laughter, sighs and coughs.

Each row pairs a real recording with a synthetic recording in which the edit has been made. It also gives a natural-language instruction for each direction. Both recordings are embedded in the Parquet files, so every row is self-contained. Audio samples are on the demo page.

Contents

• Overview
• Statistics
• Usage
• Fields
• Audio encoding
• Construction
• Benchmark results
• Refined Emilia transcripts
• Responsible use
• License
• Citation

Overview

Directories and edit_type follow the editing direction used in the paper and by our models (see Direction under Fields). The example instructions below are taken from the data and are given in that direction.

Config	What is edited	Example instructions	Pairs
single_span	One span is inserted, deleted or substituted.	Replace 'maimed them' with 'killed them'. 在‘看看’后面加上‘亮度’	29,983,217
multi_span	Two to six edits; each is an insertion, deletion or substitution.	Replace 'They' with 'We'. Add 'into a ball' after 'mud'. 在‘西’前面加上‘得悉到底有什么绝世’，在‘色’后面加上‘难看’	9,997,856
location_reference	One edit, located by occurrence ("every", "the second"), sentence boundary or timestamp.	Delete every 'greek'. Add "Actually" before the sentence-initial "Everyone". 将第2.2至3.1秒的内容替换成‘唱片厂牌’	1,199,999
nonverbal	Nonverbal events are inserted (reverse_instruction) or deleted (forward_instruction); each row holds both.	Add '[laugh]' after 'that in acting'. 删除‘想起这些’后面的‘[cough]’	210,376 × 2

• location_reference has four locator types: all_occurrences, nth_occurrence, sentence_position and timestamp. Each type has about 150,000 pairs per language, 50,000 for each edit type.
• nonverbal stores each insertion / deletion pair once, since both directions use the same two recordings, in nonverbal/<language>/insertion/. Its 210,376 rows are 210,376 insertion pairs and 210,376 deletion pairs, the paper's 210K + 210K: read a row from synthetic to original with reverse_instruction for insertion, and from original to synthetic with forward_instruction for deletion. The Pairs column of the Statistics table counts rows.
• mini (the default config) holds the first 1,000 pairs of every config / language / edit type, for previewing. Rows are shuffled, so these pairs are a random sample.

Statistics

Config	Language	Edit type	Pairs	Original (h)	Synthetic (h)	Size (GB)
single_span	en	deletion	4,996,287	11,612	12,334	1,393.5
single_span	en	insertion	4,996,777	11,726	10,419	1,238.8
single_span	en	substitution	4,998,537	11,294	11,169	1,347.3
single_span	zh	deletion	4,995,986	11,452	12,555	1,396.5
single_span	zh	insertion	4,998,088	10,143	8,663	1,047.6
single_span	zh	substitution	4,997,542	10,641	10,636	1,229.8
location_reference	en	deletion	200,000	661	601	70.1
location_reference	en	insertion	200,000	604	629	72.0
location_reference	en	substitution	200,000	602	597	68.8
location_reference	zh	deletion	200,000	681	621	69.8
location_reference	zh	insertion	199,999	621	653	72.2
location_reference	zh	substitution	200,000	633	633	70.2
multi_span	en	multi	4,998,517	12,395	12,039	1,421.8
multi_span	zh	multi	4,999,339	11,821	11,900	1,339.5
nonverbal	en	insertion	3,589	8	7	1.8
nonverbal	zh	insertion	206,787	578	523	74.8
Total			41,391,448	95,471	93,978	10,914.3

• 2,900,661 of the 11,464,043 32-bit float recordings peaked above full scale and were scaled down to a peak of 1.0 before 16-bit FLAC encoding (see Audio encoding).
• 2,989 pairs whose original Emilia recording appears in our evaluation benchmarks were removed.

Distributions (see Figure 2 of the paper):

• Duration. Utterances last at most about 30 s. The median is 6.7 s for Chinese and 7.1 s for English.
• Multi-span edits per instruction. The shares of k = 2, 3, 4, 5 and 6 are 35.4 / 32.4 / 19.2 / 9.2 / 3.9 % for Chinese and 35.5 / 30.1 / 19.4 / 10.3 / 4.7 % for English.
• Edit length. Edited spans are short. Depending on the edit type, the median is 2–4 characters in Chinese and 1–3 words in English.
• Nonverbal events. There are 224,341 events in 39 classes. The most frequent are breath (29.7 %), laughter (13.5 %) and sigh (9.3 %). 98.1 % of the events are in Chinese.

Usage

Files are organized as data/<config>/<language>/<edit_type>/. Language is zh or en. Edit type is one of deletion, insertion, substitution or multi.

from datasets import load_dataset

ds = load_dataset("zkniu/VoxCanvas", "single_span", split="train", streaming=True)
# A finer subset, selected by path:
ds = load_dataset("zkniu/VoxCanvas", data_files="data/single_span/zh/insertion/*.parquet", split="train", streaming=True)

row = next(iter(ds))
# The paper's direction for single_span: edit the synthetic recording back into the real one (see Direction).
source, target, instruction = row["synthetic_audio"], row["original_audio"], row["reverse_instruction"]

Fields

Field	Description
id	<config>/<language>/<edit_type>/<key>; unique. For Emilia-based rows the key contains the Emilia utterance id.
task, language, edit_type	Config, language and edit type. edit_type is the edit in the paper's direction (see Direction below).
original_audio	The real recording.
synthetic_audio	The edited recording, synthesized.
original_text, synthetic_text	Transcripts of the two recordings. For nonverbal, events are written as tags such as [breath].
original_duration, synthetic_duration	Durations in seconds.
forward_instruction	Instruction that turns original_audio into synthetic_audio.
reverse_instruction	Instruction that turns synthetic_audio back into original_audio. Null for location_reference.
forward_augmented_instructions, reverse_augmented_instructions	One paraphrase of each instruction, made by filling a randomly chosen template of the same kind (see Instruction paraphrases). Empty for location_reference and nonverbal.
edit_regions	Time spans (seconds) of the edits on both recordings: original_start/end and synthetic_start/end. Null when the synthesis log has no spans.
num_edits	Number of edits (for nonverbal, the number of nonverbal events).
extra	JSON string with config-specific fields; see below.

Direction. The instruction fields are tied to the recordings: forward_instruction turns the real recording (original_audio) into the synthetic one, and reverse_instruction does the opposite. Directories, edit_type and id follow the direction used in the paper and by our models:

Task in the paper	Directory	Input → target	Instruction
Single-span insertion	single_span/*/insertion	synthetic → original	reverse_instruction
Single-span deletion	single_span/*/deletion	synthetic → original	reverse_instruction
Single-span substitution	single_span/*/substitution	synthetic → original	reverse_instruction
Multi-span	multi_span/*/multi	synthetic → original	reverse_instruction
Location-reference	location_reference/*/<edit type>	original → synthetic	forward_instruction
Paralinguistic insertion	nonverbal/*/insertion	synthetic → original	reverse_instruction
Paralinguistic deletion	nonverbal/*/insertion	original → synthetic	forward_instruction

For example, the synthetic recording of a single_span/*/insertion row lacks a few words; its reverse_instruction adds them back, which gives the real recording. Using real recordings as targets gave higher speaker similarity and UTMOS than training on synthetic targets (paper, Section 5.2.5). location_reference keeps the forward direction because location-dependent instructions cannot always be inverted reliably.

Instruction paraphrases. Each instruction of single_span and multi_span is parsed into its kind and slots (the text to delete or replace, the new text, the anchor) and rewritten with one of 10 templates per kind and language, chosen at random. The kinds are replacement, deletion without an anchor, deletion before / after an anchor, and insertion before / after an anchor; an insertion whose anchor is the first (last) word of the input becomes an insertion at the beginning (end) of the speech. A multi-span instruction is paraphrased clause by clause, in the same order. Deletions keep an anchor only where the original instruction has one, i.e. where the deleted text occurs more than once in the input.

extra (json.loads(row["extra"])):

Config	Keys
single_span	align_word (word timestamps of the original recording), fix_duration (seconds of each synthesized span)
location_reference	subset, locator (all_occurrences, nth_occurrence, sentence_position, timestamp), align_word, edited_asr_text (ASR of the synthetic recording)
multi_span	edit_types; sub_edits (one entry per edit, ordered by text position: op, target_word, new_word, direction, region (character span in original_text), left_anchor, right_anchor, instruction, reverse_instruction); edited_asr_text. op, edit_types and instruction describe the original → synthetic direction; reverse_instruction is the edit's clause in the row's reverse_instruction.
nonverbal	source_dataset; nv_tag; nv_segments (removed segments on the original timeline: start, end, fix_dur, status, tag); alignment; position-free instructions delete_instruction / add_instruction

Audio encoding

• Real recordings. Emilia MP3 files are stored byte-for-byte; their sampling rates are 24, 32 or 44.1 kHz. The original recordings of nonverbal keep their source format: MP3 is stored as is, and WAV is stored as lossless FLAC.
• Synthetic recordings. These are 24 kHz mono FLAC.
  • Synthesized as 16-bit PCM: bit-exact.
  • Synthesized as 32-bit float: quantized to 16 bit, with error ≤ 1 LSB. A float recording that peaked above full scale was first scaled down to a peak of 1.0, so nothing is clipped. Its absolute level is therefore lower than the original synthesis.
• No resampling. To get a single rate, use ds.cast_column("original_audio", Audio(sampling_rate=24000)).

Construction

1. Source speech. Content editing uses the Chinese and English subsets of Emilia. Paralinguistic editing uses six nonverbal speech corpora: Emilia-NV, NonverbalSpeech-38K, SynParaSpeech, NonverbalTTS, SMIIP-NV and MNV-17. Duplicate nonverbal clips were removed by audio fingerprinting.
2. ASR and forced alignment. All source audio was re-transcribed with Qwen3-ASR-1.7B and aligned at the word level with Qwen3-ForcedAligner-0.6B.
3. Instructions.
   • The text to edit was selected by predefined rules.
   • For insertion and substitution, Qwen3.6-35B-A3B generated the new content from the transcript and the selected target, keeping the edited transcript fluent and coherent. Deletion removes the selected content directly.
   • Instructions were assembled from templates.
   • For multi-span editing, the LLM generated two to six non-overlapping edits in a single pass.
   • For paralinguistic editing, event positions come from the corpus annotations and the timestamps of the surrounding words.
4. Synthesis.
   • F5-TTS (F5TTS_v1_Base) regenerated the edited regions, conditioned on the edited transcript and the surrounding audio. Each masked region was extended by 0.5 s on both sides to avoid audible discontinuities.
   • The mel spectrogram outside the masked regions is copied from the original. The whole utterance is then vocoded with Vocos, so the waveform also differs slightly outside edit_regions.
5. Quality control.
   • Synthetic recordings longer than 30 s were removed.
   • The synthetic recording was transcribed with Qwen3-ASR-1.7B and kept only if the transcript matched the expected one exactly. English requires a word error rate of 0. Chinese requires a pinyin error rate of 0, so homophone characters in the ASR output are not penalized. For nonverbal editing, the expected transcript excludes the nonverbal events.
   • The speaker similarity between the original and synthetic recordings (WavLM-TDNN embeddings) had to exceed 0.5.
   • Human annotators spot-checked random samples of the remaining pairs.
6. Selection.
   • single_span keeps, per language and edit type, the 5M pairs whose synthetic recording has the highest UTMOS.
   • multi_span keeps the 5M pairs with the highest speaker similarity per language.
   • location_reference keeps the 50K pairs with the highest speaker similarity per language and subtype.
   • Pairs whose original Emilia recording appears in our evaluation benchmarks were removed.

Benchmark results

VoxCanvas-Edit is AuK fine-tuned on VoxCanvas-41M with LoRA (rank 16; about 20M trainable parameters, about 1.3 % of the AuK backbone) for one epoch. All results are averaged over three seeds. "AuK + GigaEdit-L" is fine-tuned in the same way on GigaEdit-L, which contains English data only.

Benchmark	Metric	AuK	AuK + GigaEdit-L	VoxCanvas-Edit
SpeechEditBench, content (zh / en / avg.)	ACC ↑	88.32 / 93.78 / 91.05	– / 94.22 / –	93.66 / 96.78 / 95.22
SpeechEditBench, paralinguistic	ACC ↑	53.25	–	54.50
Ming-Freeform-Audio-Edit, semantic (Basic-zh / Full-zh / Basic-en / Full-en)	ACC ↑	92.03 / 91.05 / 85.93 / 85.69	– / – / 83.94 / 83.50	92.62 / 91.58 / 87.51 / 86.84
Ming-Freeform-Audio-Edit, semantic (same order)	WER ↓	4.84 / 3.09 / 4.05 / 4.11	– / – / 5.80 / 5.95	4.40 / 2.93 / 3.68 / 3.77

The paper also uses the dataset to study the learning dynamics of speech editing:

1. Deletion converges fastest, then substitution, then insertion. More data mainly helps insertion and substitution.
2. Task interactions depend on the task mix. Adding substitution largely removes the slowdown of insertion seen when insertion and deletion are trained together.
3. A balanced mix of insertion, deletion and substitution gave the best trade-off between convergence and final performance among the mixtures tested.
4. Editing can be viewed as an understanding-grounded generation task that helps speech generation: joint training with TTS improves generation learning efficiency at comparable quality, and also speeds up editing convergence.
5. Initializing from a TTS model speeds up editing convergence, especially for insertion and substitution.
6. A curriculum from single-span to multi-span editing helps multi-span editing.
7. Reversing the pairs, so that real recordings are the targets, reduces the quality loss caused by synthetic targets.
8. Paralinguistic insertion is learned faster than content insertion. Both deletion abilities follow similar trajectories once the generated speech is intelligible.

Refined Emilia transcripts

The source Emilia audio used to build VoxCanvas was re-transcribed with Qwen3-ASR-1.7B and aligned at the word (English) / character (Chinese) level with Qwen3-ForcedAligner-0.6B. We release those transcripts and alignments here as standalone text data. The appendix of the paper reports the gain from training a TTS model on these re-transcribed texts instead of the original Emilia transcripts (lower WER).

data/refine_emilia_text/{en,zh}/<LANG>_B*.jsonl holds one JSON object per line with id (the original Emilia utterance id), wav, asr_text, align_word ({text, start, end} in seconds) and duration. No audio is included; join on id, not on line order.

Download only this folder, without the audio configs:

hf download zkniu/VoxCanvas --repo-type dataset \
  --include "data/refine_emilia_text/**" --local-dir ./voxcanvas_text

Or stream a single language:

from datasets import load_dataset
ds = load_dataset("zkniu/VoxCanvas", data_files="data/refine_emilia_text/en/*.jsonl",
                  split="train", streaming=True)

Derived from Emilia (gated, CC-BY-NC); obtain upstream access separately. The repository LICENSE applies.

Responsible use

The synthetic recordings imitate the voices of the speakers in the source corpora. Do not use this dataset to impersonate real people or to deceive.

License

CC BY-NC-SA 4.0, for non-commercial research. The source corpora have their own licenses and terms, which also apply. See LICENSE for those terms, the attribution notices and the citations.

Citation

@misc{niu2026voxcanvas,
  title         = {VoxCanvas-41M: A Large-Scale Bilingual Dataset for Instruction-Based Speech Editing and Its Learning Dynamics},
  author        = {Niu, Zhikang and Tu, Wenming and Li, Yalin and Huo, Yanru and Liu, Junxi and Wang, Tianrui and Yan, Ruiqi and Liang, Junxuan and Zheng, Zilong and Ma, Ziyang and Chen, Xie},
  year          = {2026},
  eprint        = {2611.08251},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2611.08251}
}