Datasets:
Chatterbox Hindi exaggeration experiment
A seeded experiment of 50,000 planned requests with individually shuffled exaggeration settings spanning 0.0–1.0, including both endpoints. 49,997 waveforms were saved (about 121.27 hours): 49,891 EOS-complete and 106 token-limit outputs. Three requests remain unresolved and are explicitly preserved. The earlier five-sample randomized pilot is available separately. The superseded 30k experiment is not part of this release.
The repository is public with manual access approval. Request access using the Hugging Face dataset page; the owner approves requests. Public metadata does not imply unrestricted rights to the underlying reference speech.
Contents and joins
| Config | Splits and rows | Contents |
|---|---|---|
samples |
train 49,891; flagged 106; pilot 5 | Original PCM16 WAV bytes, codec tokens, raw T3 tokens, text, reference key, exact parameters and generation metadata |
tokens |
same as samples | Same rows without waveform bytes; convenient for token experiments |
references |
train 18,143; pilot 5 | Original cropped WAV bytes, speaker, source dataset/shard/row/filename, crop timing, transcript and SHA256 |
requests |
train 50,000; pilot 5 | Every immutable planned request, parameters, reference mapping and saved/unresolved status |
unresolved |
train 3 | Requests without saved waveforms, errors and attempt counts |
diagnostics |
train | Failure events and available raw failed T3 tokens; never counted as successful output |
metadata |
train | JSON contents of provenance, design, audit and validation reports, stored as strings inside Parquet |
transcripts |
train | Original supplied transcript input rows used to construct the independently sampled text pool |
source_pool |
train | Eligible original reference pool metadata, without source waveforms |
Join samples/requests/tokens to references using reference_key. Join sample rows across configs using example_id. Both include the experiment namespace; pilot IDs intentionally overlap local main IDs. sample_id and reference_id alone are not globally unique. All audio is embedded in Parquet as {bytes, path}, with the original WAV bytes preserved. The path is a portable filename, not a link to a local server. A reference is embedded once per experiment and joined to every request that uses it.
Each sample contains exaggeration, cfg_weight, temperature, language_id, seed, decoder_seed, max_new_tokens, top_p, min_p, repetition_penalty, speaker_id, text, grid_index, text_slot_id, reference_key, output/token hashes, quality flags and available execution details. Parameters use float64 to preserve the recorded values. Main records also contain batch/run IDs, sizes, sampler, precision and retry metadata when originally recorded. Missing pilot execution fields are null rather than guessed.
Experiment design
- Main seed: 20261015. Exaggeration is
numpy.linspace(0, 1, 50000)shuffled before speaker/reference/text draws. There are 50,000 unique requested values, at a step of1/49999, with 5,000 requests per tenth of the range. Saved outputs retain 4,999 values in the 0.2–0.3, 0.5–0.6 and 0.6–0.7 bins; every other tenth has 5,000. - Speakers are drawn uniformly from 367 speakers, followed by an eligible recording uniformly within the speaker. References and texts are sampled with replacement. Text selection is independent of the exaggeration permutation. Crops are cached per chosen original recording. This is a randomized marginal sweep, not a full factorial experiment with every text/reference repeated at every value.
- Fixed main settings:
cfg_weight=0.5,temperature=0.5,language_id="hi",max_new_tokens=1000,top_p=0.95,min_p=0.05,repetition_penalty=1.2. - Main sampler: per-sample NumPy PCG64 uniforms and inverse CDF; decoder seed = recorded token seed + 1,000,000. Precision: float32. Generation used A100 and L40S GPUs, with batch sizes 32 and 16, and singleton retries. Exact run source snapshots and dependency versions are included.
- Pilot: five values
{0, 0.25, 0.5, 0.75, 1}, shuffled among five distinct speakers; official sequentialtorch.multinomialgeneration. Only the original overall seed was recorded. Itsdecoder_seedis null; the main decoder-seed rule does not apply to the pilot.
Audio and token interpretation
Generated WAV: mono, 24 kHz PCM16, watermarked by the model. Reference WAVs retain their original crop sample rates (recorded per row). s3_tokens are valid int32 codec IDs at 25 Hz, in [0, 6560]. raw_t3_tokens also preserves special tokens (SOS 6561, EOS 6562); raw_t3_original_shape allows the original array shape to be restored. The final token's audio is trimmed by the inference pipeline, so saved waveform frames equal (token_count - 1) * 960. NPY-file hashes refer to the original saved arrays, not Parquet serialization.
Quality and limitations
train is a loading convenience, not an evaluation split. EOS-complete does not mean transcription or listening quality has been reviewed. Keep flagged outputs separate: they hit the token limit and may contain repetition or truncation. Per-row clipping measurements and review flags are preserved. Three unresolved requests (sample_011037, sample_012050, sample_020880) remain visible with diagnostics; they are not included as saved audio. No listening or ASR review was performed for this release. For evaluation, reserve unseen speakers and texts to avoid leakage.
Every exported sample was checked against the immutable plan for parameters, text, speaker and reference mapping. All reference/generated WAV and original filtered-token hashes were checked. Raw-token filtering, EOS status, audio duration and every Parquet readback were checked. See randomized_50k/metadata/export_validation.json, coverage_and_mapping_audit.json and parquet_inventory.json. The archived design.json records the original planning phase (references_materialized: false); the exported reference rows and final validation establish that all 18,143 selected clips were subsequently materialized.
Load
from datasets import load_dataset, Audio
repo = "routsourav1729/chatterbox-experiment"
samples = load_dataset(repo, "samples", split="train", streaming=True, token=True)
references = load_dataset(repo, "references", split="train", streaming=True, token=True)
requests = load_dataset(repo, "requests", split="train", streaming=True, token=True)
# To inspect embedded bytes without loading an audio decoder:
samples = samples.cast_column("audio", Audio(decode=False))
first = next(iter(samples))
print(first["example_id"], first["reference_key"], first["exaggeration"])
Use a dataset commit SHA as revision= for a stable release. The token-only config avoids downloading waveform bytes. The unresolved and diagnostic configs permit inspecting failures without mixing them into successful output.
Reproduce
See randomized_50k/reproduction/README.md. It includes portable restoration and replay tools, exact generation source snapshots for all three runs, pinned model revisions with SHA256 checks, vendored inference code, supplied transcript rows, eligible reference metadata and recorded environments. Models are downloaded from the original repositories rather than redistributed here. The archived plan SHA256 is e4eb671d7db526190bcc42e8f860438739e7a5c4dc4105ffc30ba87a94d1fe5a.
Restoring the saved inputs, labels, token arrays and output bytes is exact. Regenerating waveforms may differ across GPU hardware, batch composition, kernels or package builds; source snapshots and seeds do not guarantee bitwise GPU reproduction. The pilot uses a different sampler, as documented above.
Provenance and rights
Reference speech comes from SayantanJoker/original_IndicVoices_R_Hindi_ at revision 648b456060c1cde9df99649e800fbe8c1280c8d1. This release preserves the actual local processed crops used for generation. Original source shard, row, filename, speaker, crop and hashes are retained; original remote audio bytes were not independently compared to the local processed source. The upstream source dataset card did not declare a license at packaging time; no blanket open license is asserted for the reference audio or derived collection. See LICENSE.md.
Model: ResembleAI/Chatterbox-Multilingual-hi, revision 82ca71273cc2a9ab19efdf8315f865c1a5af0ee7, with base voice/tokenizer assets from ResembleAI/chatterbox, revision 5bb1f6ee58e50c3b8d408bc82a6d3740c2db6e18. Vendored inference source is pinned to 449843a7ac17ba81773f2880577142e0c5f778c3; original notices are retained.
- Downloads last month
- 433