Datasets:

viai957
/

indic-duplex-speech-artifacts

an Indic SentencePiece tokenizer;
measurements of the Mimi neural codec on Indic speech;
tokenizer fertility measurements;
the utterance arrangement and split labels used in the project's duplex experiments;
a seed set of synthetic two-speaker Hindi dialogues.

The matching corpus of Mimi codec tokens from open Indic speech datasets is published separately as viai957/indic-duplex-speech-mimi-tokens. This repository contains no third-party corpus audio.

TL;DR


Configs	`codec_benchmark`, `tokenizer_fertility`, `platform_v2_arrangements`, `synth_duplex`, `backchannel_bank`
Files	65 files, 99 MB
Rows	314 codec measurements · 1,300 fertility rows · 37,046 arrangement rows · 20 dialogues · 10 backchannel clips
Languages	13 Indic languages in the measurements; Hindi (with Hinglish code-mixing) in the arrangements and dialogues
Audio	Synthetic only: 20 dialogues, 0.28 h, two channels at 24 kHz WAV, plus 10 backchannel clips
Tokenizer	SentencePiece Unigram, 65,536 pieces, 13 Indic languages plus English
Licence	CC BY 4.0. Upstream attributions are listed below.
Status	Research release. The synthetic dialogues failed the project's own quality gate (see below).

Load

from datasets import load_dataset

codec = load_dataset("viai957/indic-duplex-speech-artifacts", "codec_benchmark", split="test")
fert  = load_dataset("viai957/indic-duplex-speech-artifacts", "tokenizer_fertility", split="test")
arr   = load_dataset("viai957/indic-duplex-speech-artifacts", "platform_v2_arrangements", split="train")

# Tokenizer
from huggingface_hub import hf_hub_download
import sentencepiece as spm
sp = spm.SentencePieceProcessor(model_file=hf_hub_download(
    "viai957/indic-duplex-speech-artifacts", "tokenizer/omnivoxtral_sp.model", repo_type="dataset"))
print(sp.encode("आज कुछ नया बनाने का मन है", out_type=str))

# Synthetic dialogue audio
wav = hf_hub_download("viai957/indic-duplex-speech-artifacts",
                      "synth_duplex/audio/wsb-c0ffee-000-p13.wav", repo_type="dataset")

Sources and licences

Material in this repo	Derived from	Upstream licence
`codec_benchmark` (measurements only; no audio)	utterances sampled from google/fleurs and ai4bharat/IndicVoices; codec kyutai/mimi	CC BY 4.0
`tokenizer_fertility`, `tokenizer/`	FLEURS transcripts	CC BY 4.0
`platform_v2_arrangements` (ids and labels; no text)	IndicVoices Hindi and ai4bharat/Shrutilipi Hindi	CC BY 4.0
`synth_duplex`, `backchannel_bank`	scripts written for this project; speech rendered with maya-research/Veena	Veena model: Apache-2.0

This repository is released under CC BY 4.0. When you use it, please also credit FLEURS (Conneau et al., 2022), IndicVoices (Javed et al., 2024) and Shrutilipi (Bhogale et al., 2023).

Configs in detail

`platform_v2_arrangements`

These labels define the Hindi utterance pool and the splits used in the project's duplex experiments. Hours are measured from the Mimi codes (frames / 12.5 Hz), not from the original audio.

Corpus	train rows	train h	val rows	val h	heldout rows	heldout h
IndicVoices Hindi (`iv_gold_hi`)	13,299	23.1	2,726	4.7	2,743	4.7
Shrutilipi Hindi (`shrutilipi_hi`)	13,048	18.8	2,588	3.7	2,642	3.8
Total	26,347	41.9	5,314	8.5	5,385	8.5

Median utterance length is 5.2 s for IndicVoices and 4.4 s for Shrutilipi. Median token count is 546 and 483 tokens (95th percentile: 1,554 and 1,134), counting 21 tokens per 0.2 s window (1 text token plus 20 Mimi tokens).

Field	Type	Example
`utt_id`	int64	`279951824556237`
`corpus`	string	`shrutilipi_hi`
`arrangement`	string	`follow` (the only value in this release)
`split`	string	`train` / `val` / `heldout`
`npy`	string	`shrutilipi/hin/hi/hin_hindi_train-00003-of-00196_000014_bbd22ca0.npy`

npy points into the token corpus layout. Its file stem equals the id column of viai957/indic-duplex-speech-mimi-tokens, so the two can be joined. Transcripts were left out on purpose; the token corpus carries them.

`tokenizer_fertility` and `tokenizer/`

The parquet holds per-sentence measurements for Mistral's 32k BPE on 100 FLEURS sentences per language. The OmniVoxtral numbers come from tokenizer/comparison.md. On the same text, mean fertility falls from 10.79 tokens per word to 1.91.

Language	Mistral BPE	OmniVoxtral SP	Language	Mistral BPE	OmniVoxtral SP
Malayalam	22.20	2.49	Assamese	8.73	1.81
Odia	18.19	2.05	Bengali	7.39	1.70
Telugu	12.72	2.11	Marathi	7.34	1.83
Punjabi	12.56	1.77	Nepali	7.14	1.85
Gujarati	12.33	1.86	Hindi	5.34	1.51
Kannada	11.51	2.13	Urdu	4.58	1.56
Tamil	10.24	2.16

Field	Type	Example
`tokens`, `words`, `chars`	int	`185`, `21`, `140`
`fertility`	float	`8.81` (tokens per word)
`chars_per_token`	float	`0.757`
`unk_count`, `unk_rate`	int, float	`0`, `0.0`
`language`, `language_name`	string	`asm`, `Assamese`
`text_preview`	string	first 80 characters of the sentence

The tokenizer files are omnivoxtral_sp.model and omnivoxtral_sp.vocab: a 65,536-piece Unigram model trained on 27,606 sentences.

`codec_benchmark`

Quantizers	Utterances	Mean PESQ	Mean F0 correlation	Mean SI-SNR (dB)
4	157	1.46	0.740	-2.80
8	157	2.00	0.779	3.89

Columns: file, duration_sec, num_quantizers, pesq, si_snr_db, f0_correlation, energy_correlation, speaking_rate_ratio, encode_ms, decode_ms, encode_rtf, decode_rtf, language.

PESQ was built for waveform codecs. A neural codec resynthesises audio, so its PESQ is low even when it sounds good. The project accepted a codec at PESQ 2.0 instead of the usual 3.0, and treated F0 correlation as the more reliable signal.

The chart shows the 13 languages with 10 utterances each. Nine more codes (brx, doi, kas, kok, mai, mni, san, sat, snd) have 3 rows each with identical scores. Those rows are placeholders, not measurements.

`synth_duplex` and `backchannel_bank`

As far as the project could find, no open two-channel conversational corpus exists for Indic languages. These 20 dialogues test a synthetic route to one.


Dialogues	20 (0.28 h total; 41.7 to 62.0 s each, median 53.3 s)
Planned events	106 turns, 20 interruptions, 34 backchannels
Audio	2-channel WAV at 24 kHz, one speaker per channel, two fixed Veena voices per dialogue
Backchannel bank	10 clips: `haan`, `hmm`, `accha`, `arre`, `oho`, `wah`

synth_duplex/metadata.csv has the columns file_name, dialog_id, bytes and meta_json. meta_json holds the full timeline: a list of planned events, each with kind, speaker, onset_s and duration_s; interruptions add interruption_type and overlap_s. scripts/*.json hold the text of each turn, for example:

{"speaker": 0, "text": "आज कुछ नया बनाने का मन है, कोई recipe बता।", "markers": []}

Processing pipeline

flowchart LR
  F[FLEURS]
  IV[IndicVoices]
  SH[Shrutilipi]
  F -->|27,606 sentences| T[Train SentencePiece Unigram 65k]
  T --> FT[Fertility vs Mistral BPE<br/>100 sentences / language]
  F & IV -->|10 utt / language| C[Mimi encode/decode<br/>4 and 8 quantizers]
  C --> CB[codec_benchmark<br/>PESQ, SI-SNR, F0, energy, RTF]
  IV & SH -->|Hindi utterances| TOK[Mimi tokens + MMS-FA text alignment<br/>stride 21 interleave]
  TOK --> ARR[platform_v2_arrangements<br/>follow arrangement, train/val/heldout]
  S[Hand-written Hindi/Hinglish scripts<br/>with interruption markers] --> V[Veena TTS, 2 voices]
  V --> TL[Timeline: turns, interruptions,<br/>Poisson backchannels, calibrated offsets]
  TL --> SD[synth_duplex 2-channel WAV]
  SD --> G{Quality gate<br/>WER, speaker cos, bleed, overlap JSD}
  G -->|0 / 20 pass| Q[released as worked example]

Quality filters and decontamination

Splits: train, val and heldout in platform_v2_arrangements are utterance-level holdouts. No utterance appears in more than one split (37,046 unique npy paths). They are not guaranteed to be speaker-disjoint.
Upstream evaluation splits: all 18,768 IndicVoices utterances come from the upstream train split. No upstream valid audio is included.
Codec placeholders: the nine 3-row language codes are flagged above as not usable.
Synthetic dialogue gate (synth_duplex/m4_scores.json): the thresholds were ASR WER ≤ 0.25 per channel, speaker cosine ≥ 0.65, channel bleed ≤ -30 dB, and overlap and gap distribution JSD within sanity bounds. 0 of 20 dialogues passed. The main failure is ASR WER (dialogue 000 scored 0.66 mean WER across channels). All 20 are tagged as quarantined.
Credentials: a pattern scan for credentials and tokens across all files found none.

How the data was used

These artifacts supported two experiment tracks in OmniVoxtral on 4× NVIDIA L4. The numbers below come from the project's experiment traces.

PT-0 duplex format probe (platform_v2_arrangements): a LoRA adapter on a pretrained duplex backbone, trained on the follow arrangement. In a free-run gate on 12 held-out clips, the response score went from 0.000 (baseline) to 0.059 (adapter), against a teacher-forced 0.967. The target was about 0.62, so the gate failed, but the mechanism did fire on every clip. Both arms looped on filler such as "जी जी जी".
From-scratch audio-only scaling ladder (Mimi tokens from the companion repo): a random-init summed-stream model on a 5-hour Indic slice, trained for 6,000 matched steps. Validation cross-entropy was 10.05 at 14.0 M parameters, 9.92 at 20.0 M and 9.86 at 31.6 M. The context-permutation gap was 3.37, 3.64 and 3.46 nats. That is directionally better with scale, but there was one seed and no confidence intervals, so it is not a scaling law.
Codec and tokenizer choice: codec_benchmark set the 8-quantizer Mimi configuration and its acceptance threshold, and the fertility study motivated the 65k Indic tokenizer.

No model trained on this data has been released publicly.

Intended use

Reference measurements for choosing a codec and tokenizer for Indic speech models.
A reusable Indic SentencePiece tokenizer.
Reproducible split definitions over IndicVoices and Shrutilipi Hindi.
A worked example of synthetic full-duplex dialogue generation.

Limitations

Codec and fertility numbers come from small samples (10 utterances and 100 sentences per language) of FLEURS-style read speech. Conversational and code-mixed speech will behave differently.
The tokenizer covers 13 Indic languages plus English, not all 22 scheduled languages.
The arrangement labels cover Hindi only and use a single arrangement type.
The synthetic dialogues are few, Hindi only, TTS-read, and below the project's quality bar. Do not train on them as if they were real conversations.

Ethics and personal data

The repository holds no recordings of real people. The synthetic voices come from an open TTS model, and the scripts are fictional. Arrangement rows are opaque utterance ids from corpora that their publishers released under CC BY 4.0, with participant consent recorded by those publishers. text_preview holds FLEURS sentences, which are Wikipedia-derived read text.

Citation

@misc{indic_duplex_speech_artifacts_2026,
  title        = {Indic Duplex Speech Artifacts},
  author       = {Vignesh},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/viai957/indic-duplex-speech-artifacts}}
}