Vietnamese Transcription: Six Tones, Real Accuracy

Getting a clean Vietnamese transcript means getting the tone marks. Drop one diacritic and you change the word: "ma" (ghost), "má" (mother), "mà" (but), "mả" (tomb), "mã" (horse), "mạ" (rice seedling) are six different Vietnamese words distinguished entirely by the tone mark on the same syllable. This is not a formatting issue, it is a meaning issue, and it is why Vietnamese transcription is harder than it first appears for AI systems.

This guide covers the orthography, the regional dialect gap, which engines actually support Vietnamese in 2026, and what to expect from each.

The Six-Tone System and Why Diacritics Are Load-Bearing

Vietnamese is written in chữ Quốc Ngữ, a Latin-based script finalized in the early 20th century. It encodes six phonemic tones through a system of stacking diacritics: the unmarked syllable carries the level tone (ngang), and the other five tones each receive a dedicated mark.

The six tones:

• ngang (level, unmarked): "a" as in ma (ghost)
• sắc (high-rising): "á" as in má (mother)
• huyền (low-falling): "à" as in mà (but)
• hỏi (dipping-rising): "ả" as in mả (tomb)
• ngã (dipping with glottal break): "ã" as in mã (horse)
• nặng (low-falling-heavy, dot below): "ạ" as in mạ (rice seedling)

The orthography gets denser when vowel-quality diacritics are layered on top of tone marks. Vietnamese has seven modified vowels: ă, â, ê, ô, ơ, ư, and đ. A vowel like "ă" can carry any of the six tones, producing forms like ắ, ằ, ẳ, ẵ, ặ. Unicode represents these in precomposed form under Normalization Form C, which is important for transcription output: a tool that returns decomposed combining characters rather than precomposed code points can produce text that looks correct on screen but breaks in downstream text processing.

For transcription purposes, the key point is that tone marks are not optional punctuation. A transcript that returns unaccented Latin letters is not a draft, it is the wrong text.

Engine Support: Thinner Than Tier-1 Languages

Vietnamese is listed as a supported language by the major engines:

• OpenAI Whisper (including the large-v3 model): lists Vietnamese (vi) in its supported language set.
• Deepgram Nova-2 and Nova-3: both list vi with support across batch and streaming endpoints.
• AssemblyAI Universal: includes Vietnamese, with speaker diarization available.

What "supported" means varies. AssemblyAI's documentation places Vietnamese in the "Good accuracy" band, which it defines as word error rates above 10% to 25%. For comparison, English sits below 10% WER. Vietnamese is a lower-resource language in training data terms, and the honest expectation is that accuracy on Vietnamese is meaningfully lower than on French, Spanish, or German, especially for non-standard audio.

Otter.ai does not support Vietnamese. Its supported languages, as of 2026, are English, Spanish, French, German, Japanese, and Chinese (Simplified). Any tool listing Otter accuracy for Vietnamese is inventing numbers.

For a broader look at how engine support varies across languages, see why AI struggles with low-resource languages.

Northern vs. Southern vs. Central: The Training Data Gap

Vietnam has three main regional dialect groups, and AI coverage is uneven across them.

Northern Vietnamese (Hanoi Standard)

This is the dialect that dominates news media, formal education, and government content, and it represents the largest share of Vietnamese transcription training data. The six tones are clearly acoustically distinct in Northern Vietnamese. AI accuracy is highest on Hanoi-standard audio. If you are transcribing formal broadcast content from VTV or formal business speech from Hanoi-based speakers, you are working with the dialect AI was built to handle.

Southern Vietnamese (Ho Chi Minh City, Mekong Delta)

The major difference in Southern Vietnamese is the hỏi/ngã merger. In the South, both tones are realized as a simple low-dipping tone without the glottal interruption that distinguishes ngã in the North. The written language still requires both tone marks in their respective positions, but the acoustic distinction that helps an AI assign them has collapsed. Accuracy on Southern Vietnamese is generally lower than on Northern Vietnamese, particularly for distinguishing ả and ã. Research on Vietnamese ASR datasets has also documented that Southern Vietnamese is underrepresented in training corpora compared to Northern and even Central Vietnamese, which compounds the problem.

Central Vietnamese (Hue, Da Nang, Quang Tri)

Central Vietnamese is the most challenging dialect for current AI systems. It preserves consonant distinctions that Northern Vietnamese has merged, including tr/ch, s/x, and d/gi/r. Its tone realizations differ from both Northern and Southern standard forms, with dramatic pitch drops and breathy voice quality not present in the dialect AI primarily learned. The Hue dialect also retains vocabulary absent from most training data: "mô" for "đâu" (where), "răng" for "sao" (why). Budget for more post-editing on Central Vietnamese audio than on any other regional variety.

Diacritics and Orthographic Encoding: What Can Go Wrong

Because Vietnamese characters involve stacking, the Unicode representation matters in practice. A character like "ộ" (the vowel ô with the nặng dot below) is a single precomposed code point (U+1ED9) under NFC normalization. Engines that return the decomposed sequence (o + combining circumflex + combining dot below) can cause problems in editors, databases, and export pipelines that expect NFC. When evaluating an engine, test it with a known Vietnamese passage and inspect the raw output for normalization form, not just the rendered screen result.

A second practical issue: some older tools strip diacritics entirely to avoid encoding problems, returning plain ASCII output. This is not a transcript, it is a phonetic hint. Any engine you use for Vietnamese content must return diacritized output by default.

Code-Switching: Vietnamese-English in Business and Tech

Vietnamese-English code-switching is pervasive in Vietnamese tech and business speech. Sentences like "Tôi đang work on a project mới" are routine in startup environments, and the pattern is well-documented in sociolinguistics research on Vietnamese office communication and diaspora communities. The switch happens because English technical vocabulary (email, deadline, meeting, server) has no natural Vietnamese equivalent or is simply faster in context.

For transcription, this creates a dual-language problem: the engine needs to handle Vietnamese diacritics for the Vietnamese portions and standard English for the insertions, within the same utterance. Modern engines that explicitly support Vietnamese generally manage this better than English-dominant tools that try to detect Vietnamese automatically, because language detection at the utterance level frequently misfires on mixed speech.

For tips on handling mixed-language audio, see fixing multilingual code-switching in transcripts.

For older Vietnamese speakers educated before 1975, there is also Vietnamese-French code-switching, particularly in the diaspora community in France and among older residents in Saigon-era professional contexts. The volume of this content is lower, and engine support for Vietnamese-French mixing is less tested.

Practical Tips for Better Vietnamese Transcription

1. Set the language code explicitly to vi. Auto-detection on mixed-content channels sometimes defaults to Chinese or other Southeast Asian languages, especially on short audio clips.
2. Record without background noise. Vietnamese tones are realized as pitch contours spanning the full syllable. Noise degrades pitch detection directly, which degrades tone recovery. This matters more for Vietnamese than for non-tonal languages.
3. Provide a vocabulary list for names and places before you upload if your tool supports custom vocabulary. Vietnamese personal names and place names (Hà Nội, TP HCM, Đà Nẵng, Nguyễn Văn An) with correct diacritics give the engine an anchor for locally ambiguous transcription decisions.
4. Verify diacritic output immediately. Check the first paragraph of any Vietnamese transcript for missing or incorrect tone marks before you rely on the rest.
5. Expect more correction on Central Vietnamese audio. Plan for 3 to 5 additional editing passes on Hue or Da Nang speaker recordings compared to Hanoi-standard content.

For a broader guide to AI accuracy across languages and dialects, see transcription accuracy explained.

Comparing Vietnamese Transcription Tools

The table below is based on verified pricing from vendor sites as of July 2026. Accuracy is described qualitatively, not as invented percentages, because no vendor publishes Vietnamese-specific WER benchmarks for these products.

For a full breakdown of per-minute and subscription transcription costs, see transcription pricing comparison 2026. If your Vietnamese content is for a podcast or ongoing series, best transcription for podcasts 2026 covers workflow considerations.

Who Actually Uses Vietnamese Transcription

Vietnamese diaspora content creators in the US (particularly California's Little Saigon community, which is the largest Vietnamese community outside Southeast Asia) produce Vietnamese-language podcasts, interview content, and family archive recordings. For diaspora audiences, diacritic fidelity matters both because the written output needs to be readable in standard Vietnamese and because dropping tone marks changes the text's meaning, not just its appearance.

Vietnamese newsrooms use transcription to accelerate interview turnaround for print and digital outlets. Vietnamese educators transcribe lectures for published course material. The common thread is that all of these use cases require diacritized output, not ASCII approximations.

For a broader view of how diarization works in multi-speaker Vietnamese content, see speaker diarization explained.

If you need a clean Vietnamese transcript and you do not need meeting-bot features, ConvertAudioToText supports Vietnamese with proper diacritized output and a free tier for testing on your own audio.

Frequently Asked Questions

Does AI transcription preserve Vietnamese tone marks?

It depends on the engine. Engines trained on properly diacritized Vietnamese text, such as Whisper and Deepgram Nova-2/3, return tone marks in their output. Engines that only support a short list of languages (Otter, for example, which does not support Vietnamese at all) will not handle it. Always verify that the tool lists Vietnamese as a supported language before you upload.

Which regional Vietnamese dialect is hardest for AI to transcribe?

Central Vietnamese, especially the Hue dialect, is the most difficult. Its tone realizations differ from the Northern standard that dominates AI training data, it retains consonant distinctions merged in the north (tr/ch, s/x, d/gi/r), and it uses vocabulary (mô, răng) absent from most training corpora. Expect noticeably more errors on Central Vietnamese than on Hanoi-standard or Ho Chi Minh City recordings.

Does the Southern Vietnamese hoi-nga merger cause transcription errors?

It can. In Southern Vietnamese, the hoi (smooth-dipping) and nga (dipping with glottal break) tones are pronounced nearly identically. An engine that processes audio phonetically may write the wrong tone mark, because the acoustic signal no longer distinguishes them. The correct written form uses both tones, so a well-trained engine should apply them based on vocabulary context, but this is a known weak spot.

How does Vietnamese-English code-switching affect transcription quality?

Modern engines generally handle it reasonably well: Vietnamese words come back diacritized in Vietnamese, English insertions come back in English. The issue is that code-switching is extremely common in Vietnamese tech and business speech, so if an engine's Vietnamese model is weak, the code-switched utterances compound the error rate. Use an engine that explicitly supports Vietnamese, not just English-dominant multilingual detection.

What file formats and audio quality matter most for Vietnamese transcription?

Format matters less than audio quality for a tonal language. Vietnamese tone marks depend on subtle pitch contours, specifically rising, falling, and glottalized shapes that span the whole syllable. Background noise and microphone quality directly degrade tone recovery. Record in a quiet environment, use a dedicated microphone if possible, and export audio at 44.1 kHz or higher. Common formats (MP3, M4A, WAV, FLAC) all work with major engines.

Sources

• Deepgram Models and Languages documentation: https://developers.deepgram.com/docs/models-languages-overview
• AssemblyAI Supported Languages documentation: https://www.assemblyai.com/docs/supported-languages
• Otter.ai Supported Languages help article: https://help.otter.ai/hc/en-us/articles/360047247414-Supported-languages
• Happy Scribe pricing page: https://www.happyscribe.com/pricing
• Sonix pricing page: https://sonix.ai/pricing/detailed-pricing-and-features
• Vietnamese phonology, Wikipedia: https://en.wikipedia.org/wiki/Vietnamese_phonology
• Vietnamese alphabet, Wikipedia: https://en.wikipedia.org/wiki/Vietnamese_alphabet
• Southern Vietnamese tones, SVFF: https://svff.online/blog/southern-vietnamese-tones-explained
• Vietnamese ASR revisit, arXiv 2025: https://arxiv.org/html/2603.14779v1
• Vietnamese-English code-switching in office contexts: https://stdjssh.scienceandtechnology.com.vn/index.php/stdjssh/article/view/690