--- license: cc-by-nc-4.0 task_categories: - text-to-speech language: - vi pretty_name: ZeroBench-TTS size_categories: - n<1K tags: - tts - zero-shot - voice-cloning - code-switching - cross-lingual - vietnamese configs: - config_name: challenging data_files: - split: test path: data/challenging.parquet - config_name: code_switch data_files: - split: test path: data/code_switch.parquet - config_name: cross_lingual data_files: - split: test path: data/cross_lingual.parquet - config_name: vietnamese data_files: - split: test path: data/vietnamese.parquet --- # ZeroBench-TTS From [ZeroWeight AI](https://huggingface.co/zeroweight-ai) — the benchmark counterpart to ZeroTTS. A small, fully specified benchmark for **zero-shot Vietnamese TTS / voice cloning**: 39 held-out Vietnamese reference voices × 3 subsets (`vietnamese`, `code_switch`, `challenging`), plus 20 non-Vietnamese reference voices × 1 `cross_lingual` subset = **137 test items**. Every item pairs one reference clip (with its ground-truth transcript) with one **Vietnamese** target sentence to synthesize — `cross_lingual` tests whether a system can speak Vietnamese given a foreign-language voice prompt, which is the direction that matters for a Vietnamese TTS system (not the reverse). It is deliberately small enough to run every checkpoint on, and every item is scored the same way — WER, SSIM, UTMOS — so numbers are comparable across systems. ## Subsets | config | n | language of `text` | what it stresses | |---|---|---|---| | `challenging` | 39 | `vi` | repeated words, acronyms (ChatGPT, ASEAN, WHO), times, dates, percentages, dense punctuation | | `code_switch` | 39 | `vi` | natural Vietnamese with embedded English (names, brands, workplace loanwords) | | `cross_lingual` | 20 | `vi` | non-Vietnamese reference voice (Emilia: zh/de/fr/ja/ko) reading Vietnamese text — many languages -> Vietnamese zero-shot cloning | | `vietnamese` | 39 | `vi` | monolingual Vietnamese — the baseline case | Within every subset the target sentences are split evenly into **short / medium / long** thirds (`length_bucket`), so length is covered across the board rather than confounded with subset. `cross_lingual`'s target texts are drawn from the other three subsets' pools (mixed, not a pool of its own) — see `text_origin_subset`. ## Fields | field | type | description | |---|---|---| | `id` | string | `{subset}/{voice_id}` | | `subset` | string | one of the configs above | | `text` | string | **the text to synthesize**, verbatim — always Vietnamese | | `text_normalized` | string | spoken-out form of `text` (equal to `text` when nothing needs expanding) | | `has_normalization` | bool | whether `text_normalized` actually differs | | `lang` | string | target language of `text` (always `vi`) | | `text_origin_subset` | string | which pool `text` came from (`vietnamese`/`code_switch`/`challenging`) — trivially equal to `subset` except for `cross_lingual` rows | | `length_bucket` | string | `short` / `medium` / `long` | | `n_chars`, `n_words` | int | length of `text` | | `ref_audio` | audio | **reference voice** to clone, mono 24000 Hz | | `ref_text` | string | ground-truth transcript of `ref_audio`, in `voice_lang` — for systems that need an in-context audio+text prompt. Speaker-encoder models can ignore it. | | `ref_duration` | float | seconds of `ref_audio` | | `voice_id` | string | stable voice identifier; the same voice appears once per subset | | `voice_source` | string | `vivos` / `viVoice` / `phoaudiobook` / `emilia` | | `source_speaker`, `source_utterance` | string | provenance in the source corpus | | `voice_lang` | string | language of the reference speaker (`vi` for the first three subsets; `zh`/`de`/`fr`/`ja`/`ko` for `cross_lingual`) | | `cross_lingual` | bool | `lang != voice_lang` — always `False` except for the `cross_lingual` subset | ## Usage ```python import io, soundfile as sf from datasets import load_dataset, Audio ds = load_dataset("zeroweight-ai/ZeroBench-TTS", "challenging", split="test") # Reading the raw wav bytes works on every `datasets` version; asking for a # decoded array needs a codec backend (`pip install torchcodec`) on datasets>=5. ds = ds.cast_column("ref_audio", Audio(decode=False)) row = ds[0] wav, sr = sf.read(io.BytesIO(row["ref_audio"]["bytes"])) audio = my_tts(text=row["text"], reference=(wav, sr)) # clone and speak ``` ## Metrics The reference implementation is `scripts/eval_tts.py` in the ZeroTTS repo. * **WER** — [`vinai/PhoWhisper-large`](https://huggingface.co/vinai/PhoWhisper-large) (a Vietnamese-specialized Whisper finetune — NOT raw `openai/whisper-large-v3`) transcribes the generated audio; WER is computed against the target text after lowercasing, punctuation stripping and whitespace collapsing. `scripts/eval_tts.py --whisper_model` can swap in any other Whisper-family checkpoint if you want a different ASR. * **SSIM** — cosine similarity between `microsoft/wavlm-base-plus-sv` x-vectors of the generated clip and `ref_audio`. * **UTMOS** — UTMOSv2 predicted naturalness MOS of the generated clip. * **Excess silence** — seconds of unwanted lead-in/tail/mid-utterance pause, specifically to catch weird long silences (stalls, dead air, a stuck decoder) that WER/SSIM/UTMOS don't penalize on their own — none of the three metrics above notice a generation that's otherwise correct but padded with several extra seconds of dead air. **Scoring the `challenging` subset:** take `min(WER(hyp, text), WER(hyp, text_normalized))`. Whisper may transcribe `3,2%` either as digits or as *ba phẩy hai phần trăm*, and that choice belongs to the ASR, not to the TTS system under test — scoring against a single form would measure Whisper's formatting policy instead of intelligibility. ## Voice selection * **VIVOS** — every speaker of the VIVOS *test* split (`VIVOSDEV01`–`VIVOSDEV19`), one clip each, chosen closest to 6 s. * **viVoice / phoaudiobook** — 10 speakers each. These corpora carry no usable speaker labels, so clips were selected geometrically: **k-means over WeSpeaker x-vectors, keeping each cluster's medoid**. That yields mutually distinct voices spread over the corpus's voice space, while medoids stay typical of their region rather than being recording outliers. `source_speaker` is therefore a synthetic cluster id, not an upstream label. All Vietnamese reference clips are 4–12 s, mono, 24000 Hz. * **Emilia (`cross_lingual` only)** — 4 speakers each from 5 non-Vietnamese languages (de, fr, ja, ko, zh), sampled from the gated [amphion/Emilia-Dataset](https://huggingface.co/datasets/amphion/Emilia-Dataset). Candidates are filtered by duration, DNSMOS quality, and a per-language text sanity check (script range for zh/ja/ko, function-word check for de/fr) — Emilia's own per-clip language tag is not fully reliable (English text was found mislabeled `ko` in the first Korean shard during a spot check). ## Provenance & license Vietnamese reference audio is redistributed from [AILAB-VNUHCM/vivos](https://huggingface.co/datasets/AILAB-VNUHCM/vivos), [capleaf/viVoice](https://huggingface.co/datasets/capleaf/viVoice) and [thivux/phoaudiobook](https://huggingface.co/datasets/thivux/phoaudiobook). `cross_lingual` reference audio is redistributed from [amphion/Emilia-Dataset](https://huggingface.co/datasets/amphion/Emilia-Dataset). Please honour each upstream corpus's own license and cite them alongside this benchmark. The target sentences are original, written for this benchmark. Released for research and evaluation use.