gavinlaw/gigaspeech_speaker_prompts_v2

GigaSpeech Speaker Prompts v2 (one-per-opus)

A pool of 9,989 single-segment English voice prompts curated from GigaSpeech for use as zero-shot voice-cloning references (CosyVoice3, XTTS, etc.).

Compared to our v1 pool (10k segments, segment-biased: at most 27 segments could come from the same podcast/youtube opus, so the effective speaker count was only ~6,206), this v2 pool enforces one segment per unique GigaSpeech opus, dramatically increasing acoustic diversity.

Contents

The repo ships as a single tarball:

• gigaspeech_speaker_prompts_v2.tar (~2.3 GB, uncompressed — wav data does not meaningfully compress)
• gigaspeech_speaker_prompts_v2.tar.sha256 — sha256 checksum

Extract to get:

gigaspeech_speaker_prompts/
├── speaker_index.json        # 9989-entry array, schema below
├── diversity_report.json     # audit report (v1 vs v2 diversity stats)
├── POD0000006662_S0000002.wav
├── YOU0000016358_S0001024.wav
└── ... (9989 total .wav, 16 kHz mono, ~6-12s each)

speaker_index.json schema

[
  {
    "speaker_id": "POD0000006662_S0000002",
    "wav_path": "/absolute/path/to/POD0000006662_S0000002.wav",  // absolute path on original machine; ignore
    "text": "uber infrastructure runs across thousands of server instances and produces terabytes of monitoring data",
    "gender": "unknown",
    "accent": "unknown",
    "domain": "podcast",
    "duration_s": 7.11
  },
  ...
]

• wav_path is the absolute path on the machine that built this pool — downstream code should reconstruct the path as <your_local_dir>/<speaker_id>.wav.
• gender / accent are "unknown" because GigaSpeech does not ship those labels; they are kept for schema compatibility with VCTK-style prompt pools.
• domain ∈ {"podcast", "youtube", "audiobook"} inherited from the GigaSpeech opus_id prefix (POD/YOU/AUD).
• duration_s is the segment length (all segments are in the 6-12 s sweet spot that CosyVoice3 likes as a reference).

Intended use

This was built to back the voice-cloning reference pool for:

• rag_tts_multispeaker_noise.py (InfiniSST) — random Random(line_idx).choice(pool) per utterance, so every synthetic utterance gets a different speaker.
• Any zero-shot TTS pipeline that takes (ref_text, ref_audio) as input.

Previous alternative was a VCTK-based pool of ~100 speakers (studio English, too narrow). This v2 pool is ~100× larger and covers wild podcast / YouTube acoustic conditions, which reduced speaker over-fitting in downstream speech-LLM training (see run 43871 in the InfiniSST WandB).

Quick start

# Download
hf download --repo-type dataset gavinlaw/gigaspeech_speaker_prompts_v2 \
    gigaspeech_speaker_prompts_v2.tar \
    --local-dir ./

# Verify
sha256sum -c gigaspeech_speaker_prompts_v2.tar.sha256

# Extract
tar -xf gigaspeech_speaker_prompts_v2.tar

Then point your TTS script at ./gigaspeech_speaker_prompts/.

Source & license

All underlying audio is extracted from GigaSpeech (Apache-2.0). This repo redistributes short 6-12 s segments under the same Apache-2.0 license. Please cite the original GigaSpeech paper if you use this pool in research:

GigaSpeech: An Evolving, Multi-domain ASR Corpus with 10,000 Hours of Transcribed Audio (Chen et al., Interspeech 2021).