Vietnamese Speech Data: Why the Region Comes Before the Hour Count
Vietnamese speakers from Hanoi, Hue and Ho Chi Minh City do not sound alike, and the tone system itself differs between north and south. Here is what that means for a speech dataset.
Vietnamese is usually described as one language with three regional accents. For a speech dataset, that description is too generous. The regional groups differ in how tones are produced, and one of them merges two tones that the others keep separate. A Vietnamese speech dataset that does not say which region it covers is underspecified.
The three groups are conventionally named for their main cities: northern around Hanoi, central around Hue, and southern around Ho Chi Minh City. Each is internally varied, and the boundary between them is not sharp.
The tone system itself differs between north and south
Standard descriptions of Vietnamese list six tones. Northern speech distinguishes all six. Southern speech merges two of them, so a pair that is contrastive in Hanoi is not contrastive in Ho Chi Minh City.
This is not an accent detail that a model can average over. A corpus that mixes regions without labelling them teaches a model contradictory tone cues, and the training loss reflects it. If the deployment target is one region, the corpus should be drawn from that region and labelled accordingly.
Central Vietnamese is the hardest region to collect and to transcribe
Central varieties, around Hue and the surrounding provinces, are the most distinct and the least represented in existing material. They are also harder to transcribe, because the written standard does not map cleanly onto how the words are pronounced.
A corpus that covers only the north and the south will systematically underrepresent central speakers. If the product will serve the whole country, that gap has to be closed deliberately, not left to whichever speakers happened to be available.
The writing system is Latin, but the diacritics carry the meaning
Vietnamese uses a Latin-based script with diacritics that mark both vowel quality and tone. Dropping the diacritics does not produce a simplified transcript; it produces a different word. Transcription has to preserve them exactly.
Annotator consistency matters more here than in most languages, because a single missing tone mark changes the word. A written convention for tone marks, and a review pass that checks them, belong in the specification.
A checklist for Vietnamese speech datasets
- Region per speaker: northern, central or southern, plus the province.
- Whether the corpus is balanced across regions or weighted toward one.
- Tone mark convention, applied consistently and reviewed.
- Reading versus spontaneous speech, since formality changes tone production.
- Channel: studio, mobile or telephone, matched to the deployment target.
- Acceptance check per region, with central speakers checked separately.
The decision that matters is which region and which tone convention, not how many hours. Both are cheap to specify and expensive to discover missing at delivery.
For related pages, see Vietnamese, the full dataset briefs, the technical reference on tone and transcription, and the conversational speech category.
Questions buyers ask
Can one Vietnamese dataset cover all three regions?
It can, but the regions have to be labelled and balanced deliberately. A corpus that pools them without recording where each speaker came from cannot be split later, and the tone systems are not fully compatible.
Why does southern Vietnamese have fewer tones?
Southern speech merges two tones that northern speech keeps distinct. The merge is systematic, not a recording artefact, and it changes which contrasts a model can learn.
Do diacritics matter in Vietnamese transcripts?
They carry meaning. Removing them does not simplify the transcript, it changes the words. Tone and vowel marks have to be preserved and checked.
Is central Vietnamese worth the extra cost?
If the product serves the whole country, yes. Central speakers are the least represented in existing corpora and the hardest to add later, because recruitment is more difficult than in the two largest cities.