EN-VI Speech 20k
English-Vietnamese speech-translation data for fine-tuning Hibiki-Zero (simultaneous speech translation) on Vietnamese→English. Built in the hibiki-zero repository.
Access: gated with automatic approval. Each folder carries its own source's terms (see Access terms); users must accept all terms that apply to the folders they access.
Folders
| Folder | Config | Rows | Source hours | Content |
|---|---|---|---|---|
data/ |
default |
696,243 | — | PhoMT text pairs, both sides synthesized |
pho/ |
pho-segments |
290,037 | 939 | Real Vietnamese from PhoAudiobook, English voiced in speaker's voice |
vivoice/ |
vivoice-segments |
269,350 | 571 | Real Vietnamese from viVoice, English voiced in speaker's voice |
giga/ |
giga-segments |
281,133 | 337 | Real Vietnamese from GigaSpeech 2 (archives 0–12), English voiced in speaker's voice |
en_speech/ |
en_speech |
3,628,994 | ~10,767 EN / ~9,477 VI | Real English read speech (Common Voice 27, Libriheavy) with Vietnamese |
The first real-Vietnamese tranche (pho + vivoice + giga: 840,520 rows) yielded 777,090 cached
rows after quality filtering (pho 98.0%, vivoice 96.6%, giga 82.7%). Their grounded Mimi + CTC
training caches are in
anquachdev/mimi-en-vi-speech-20k.
Generation
data/ — PhoMT synthetic pairs
Both sides are synthesized from VinAI PhoMT text pairs.
- Vietnamese speech: VieNeu-TTS v3-Turbo (
pnnbao-ump/VieNeu-TTS-v3-Turbo), 48 kHz mono WAV. Voice pool: 40 speakers — 12 VieNeu presets and 28 voices cloned from VIVOS speakers. Candidates were screened for transcription accuracy (PhoWhisper-small CER) and clone fidelity (speaker-embedding cosine similarity); 8 were excluded. - English speech: Kokoro-82M, 24 kHz mono WAV. Pool: 34 candidates (7 Kokoro preset voices plus same-gender blends).
- Cross-lingual timbre matching: for PhoMT index ≥ 345,600 (~51% of rows), the English voice for each row is chosen by cosine similarity of 192-d speaker embeddings to the Vietnamese voice among the 34 English candidates. Earlier rows pair voices independently.
- Assignment: voice identity and speaking speed are set per row by a seeded deterministic RNG, making the corpus reproducible from PhoMT indexes.
- Quality control: pairs whose EN/VI duration ratio falls outside 0.4–1.8 are regenerated once, then excluded if still out of range; near-silent waveforms (RMS < 1e-4) are regenerated.
pho/, vivoice/, giga/ — real-Vietnamese tranche
Follows the Hibiki-Zero paper's recipe for a new language: real source speech → transcript → English translation synthesized in the speaker's voice. Each source folder has four sub-parts:
<source>/segments/ Vietnamese audio (32 kbps Ogg Opus) + corpus transcript
<source>/text_en/ English translation of each segment
<source>/en_audio/ English speech in the speaker's voice + Qwen3-ASR transcript + WER flag
<source>/asr_vi/ Qwen3-ASR-1.7B re-transcription of the Vietnamese clip + CER
Ingest (training/tranche/ingest.py): each source shard is downloaded from its pinned
revision, its clips extracted (5–22.4 s for viVoice, whose rows are shuffled so neighbours
cannot be rejoined; PhoAudiobook and GigaSpeech 2 keep any clip length), and audio
re-encoded to 32 kbps Ogg Opus at the source rate.
| Source | Pinned revision | Rows | Hours | Speakers / channels |
|---|---|---|---|---|
| PhoAudiobook (shards 0–223) | thivux/phoaudiobook@f19f241 |
290,037 | 939 | 710 readers |
| viVoice (5–22.4 s clips) | capleaf/viVoice@693e988 |
269,350 | 571 | 186 channels |
| GigaSpeech 2 (archives 0–12) | speechcolab/gigaspeech2@8dc0d0e |
281,133 | 337 | — |
Translation (training/tranche/translate.py): Muse Spark 1.3 Contributor via OpenRouter,
in batches of up to 40 consecutive rows of a recording (providing context) and 6,000
characters. Batches that return foreign script or fail to parse are split to single rows and
retried. 840,520 rows translated with no failures.
English TTS (training/tranche/voice.py):
VoxCPM2 2B (nano-vllm-voxcpm engine; RTF ~0.007
on an H100 SXM). Each row is prompted with its own Vietnamese clip and transcript so the
output carries the source speaker's voice. Every output is verified by Qwen3-ASR-1.7B
(English forced); rows with WER > 20% or pace off by 2× are re-voiced up to 3 times. The
full run voiced 2,138 hours in 13.4 H100-hours (completed 2026-11-06). Rows whose
en_asr transcript contains Vietnamese letters absent from text_en (the model occasionally
speaks a Vietnamese clause) are flagged en_broken and excluded from the cache.
ASR re-transcription (training/tranche/filters.py): Qwen3-ASR-1.7B on CUDA, language
unforced. CER is computed over spoken text (lower-case, digits expanded to Vietnamese
words, punctuation removed). Results stored as <source>/asr_vi/.
Cache (training/tranche/cache.py): rows passing all gates are encoded into grounded
Mimi (float32, TF32 off) + wav2vec2-base-960h CTC caches and pushed to
anquachdev/mimi-en-vi-speech-20k. Gates, in order:
en_brokenflag seten_asrcontains Vietnamese letters absent fromtext_en- language detected as non-Vietnamese
- CER > 0.15
- nghìn/triệu/tỷ under-counted in transcript vs. ASR (dropped magnitude)
cache_codes's own gates: frame cap, CTC alignment score < 0.5
| Source | Rows | Kept |
|---|---|---|
| PhoAudiobook | 290,037 | 98.0% |
| viVoice | 269,350 | 96.6% |
| GigaSpeech 2 | 281,133 | 82.7% |
GigaSpeech 2's lower yield is driven by broken English TTS (11.1% of rows), concentrated in clips under 2 s where a Vietnamese name the ASR reads differently triggers the WER gate.
en_speech/ — English read-speech corpus
A file-for-file copy of
anquachdev/common-vi-en placed
here so the full training corpus sits in one repository. See that dataset's card for full
generation details. Briefly:
- English clips: Common Voice 27.0 (
cv27) and Libriheavy (libriheavy), 5–22.4 s, ≥ 10.5 kHz audio bandwidth.cv-anquach: earlier voiced Common Voice clips from that dataset's contributor, filtered and merged. - Vietnamese translations: Claude Sonnet 5.5, Gemini 3.8 Flash and Muse Spark 1.3, in batches of neighbouring lines.
- Vietnamese speech: VieNeu-TTS v3-Turbo, voice-transfer from the row's own English clip (first 8 s as the speaker reference), 48 kHz mono Ogg Opus.
| split | source | shards | rows | EN hours | VI hours |
|---|---|---|---|---|---|
| train | cv27 | 87 | 522,251 | 943.5 | 685.3 |
| train | cv-anquach | 7 | 138,918 | 256.1 | 211.6 |
| train | libriheavy | 1,745 | 2,958,303 | 9,542.2 | 8,560.6 |
| validation | cv27 | 1 | 6,021 | 11.6 | 7.8 |
| validation | libriheavy | 2 | 3,501 | 13.3 | 12.3 |
Columns
data/ (config default)
| column | type | content |
|---|---|---|
en |
string | English sentence |
vi |
string | Vietnamese sentence |
audio_en |
Audio | Kokoro-82M output, 24 kHz mono WAV |
audio_vi |
Audio | VieNeu-TTS output, 48 kHz mono WAV |
duration_en_s |
float32 | English audio length in seconds |
duration_vi_s |
float32 | Vietnamese audio length in seconds |
duration_ratio_en_vi |
float32 | EN / VI duration; kept within 0.4–1.8 |
pho/, vivoice/, giga/ segments
Configs pho-segments, vivoice-segments, giga-segments point to <source>/segments/*.parquet.
| column | type | content |
|---|---|---|
id |
string | <source>/<shard>/<index> |
source |
string | pho, vivoice or giga |
doc |
string | recording the row belongs to; a doc never spans shards |
seq |
int64 | order within the shard and doc |
speaker |
string | corpus reader or channel; empty for GigaSpeech 2 |
start_s, end_s |
float64 | clip times in the recording (trimmed audio may differ) |
text |
string | corpus transcript |
duration_s |
float64 | decoded clip length in seconds |
source_rate |
int32 | source sample rate in Hz |
audio |
Audio | 32 kbps Ogg Opus at source_rate |
The text_en/, en_audio/ and asr_vi/ sub-folders are not declared as HuggingFace configs;
their parquets share the row id with segments/.
en_speech/ (config en_speech)
| column | type | content |
|---|---|---|
id |
string | cv27/<clip>, cv-anquach/<clip>, or libriheavy/<subset>/<spk>/<book>/<rec>_<seg> |
en_audio |
Audio | English clip |
text_en |
string | cased, punctuated transcript |
text_vi |
string | Vietnamese translation |
speaker |
string | cv27/<client_id> or libriheavy/<speaker_id>; empty for cv-anquach |
en_duration_s |
float32 | English audio length in seconds |
vi_audio |
Audio | Vietnamese speech, 48 kHz Ogg Opus |
vi_duration_s |
float32 | Vietnamese audio length in seconds |
cross_lingual_timbre_matched |
bool | Vietnamese voice matches English speaker; empty until measured |
Access terms
Each folder is covered by its source's terms:
| Folder | Terms |
|---|---|
data/ |
Synthetic data, Apache-2.0; PhoMT text is VinAI's |
pho/ |
PhoAudiobook: research or education only; shared here with the author's permission; no impersonation; no speaker identification |
vivoice/ |
viVoice CC BY-NC-SA 4.0 |
giga/ |
GigaSpeech 2 access terms (SpeechColab and Tsinghua University): non-commercial research and education; clause 4 permits sharing with those who first accept the terms |
en_speech/ |
Common Voice CC0 via the Mozilla Data Collective (re-hosting prohibited; no speaker identification); Libriheavy Apache-2.0 |
Weights trained on pho/, vivoice/ or giga/ data are non-commercial (CC BY-NC-SA 4.0).
The GigaSpeech 2 terms additionally bind users to indemnification and possible access
termination.
- Downloads last month
- 33,629