sw-voice/swamiji-tts-src0-packed

Swamiji TTS - src0

One of two speaker/recording sources split out of sw-voice/swamiji-tts-merged-6h and rebuilt. 313 clips, 1.858 h, 24000 Hz mono.

Sources are kept apart because they are different recordings, and because packing across a source boundary would splice unrelated audio together.

What was changed, and why

1. Silence normalised to exactly 200 ms at each end. Trailing silence in the original corpus was bimodal - 473 clips under 50 ms against 374 over 0.8 s. In this source, 184 clips carried more than 0.8 s. That bimodality is the most likely cause of the fine-tuned model appending ~0.8 s of dead air to a one-word reply, which on a short ack is more than half the clip.

2. Consecutive clips that continue each other are packed. The source was sliced into fixed 16-22 s windows, not at utterance boundaries, so 22.7% of the original rows ended without terminal punctuation and 16.1% began mid-sentence. Training on those fragments teaches the model to start and stop mid-phrase. 46 rows here are packed (max 2 originals per row); packed_from records the originals and n_packed the count.

Result: 96.2% of clips now end on terminal punctuation, up from ~84% before.

3. Devanagari mapping refreshed. 16 transcripts changed. The corpus was inconsistently mixed - the same sentence often carried Devanagari for one term and roman for its neighbour ("Service can be done with tanu, मन or धन"), so those terms were being pronounced as English. Roman words carrying an English plural -s (ashramas, bhajans, indriyas) are deliberately left alone, since substituting a Devanagari singular would drop the plural.

Order of operations

Trim to zero silence, then pack, then re-edge to 200 ms. That order matters: if each clip kept its 200 ms before packing, every internal join in a mid-sentence continuation would carry 400 ms of silence in the middle of a phrase. Internal joins get 60 ms instead.

On splitting at "."

A sentence boundary inside an original clip cannot be cut without forced alignment, which was not available. So a packed unit is closed at the first clip whose transcript ends on terminal punctuation. Every emitted clip therefore begins at a sentence start and ends at a sentence end, and no audio is ever cut mid-clip. The residual 3.8% that do not end on punctuation are runs that hit the 30 s packing ceiling.

Shape

clips in / out	359 / 313
duration	1.858 h, mean 21.37 s, max 42.19 s
sample rate	24000 Hz mono, 16-bit
chars/s	9.29 mean

Packed rows reach ~42 s because the source windows were already ~21 s. That is within the shipped finetune budget (max_batch_tokens: 8192; a 42 s clip is ~264 audio tokens) and within our serving ceiling (MAX_GEN_LEN 300).

Sample rate is left native at 24000 Hz - VoxCPM2's data loader casts to 16 kHz itself. Resample if you consume this directly.