sw-voice/swamiji-tts-rag10-optimised

The 10 RAG-answer samples, regenerated with the current optimised config

⚠️ SYNTHETIC AUDIO — PRIVATE

Every clip is generated by a text-to-speech model. Nothing here is a recording of any person.

Unlike the public benchmark sets, these texts are actual RAG answer content about Sri Ganapathi Sachidananda Swamiji, spoken in a synthesised version of his voice. That combination is easy to mistake for a genuine recording, which is why this repo is private. Do not make it public or redistribute the clips without deciding that is intended.

Same 10 texts the original streaming ASR gate used (sw-voice/swamiji-rag-answers-10ep-v2-copy-text), so this is a like-for-like re-run of the current serving stack rather than a new benchmark.

Config

model:               sw-voice/swamiji-voxcpm2-ft-lr0.0001-10ep-deva
inference_timesteps: 10        # full quality
cfg_value:           1.5
max_model_len:       640       # NOT 400 - see below
compile_dit:         true      # torch.compile on the DiT estimator
vae_chunk:           16        # chunked VAE decode
VAE:                 bf16
GPU:                 1x RTX 3080 12 GB

Why max_model_len is 640 here and not 400

These answers are code-mixed and Devanagari costs ~1.6 tokens per character against English's ~0.21, so their prompts run 102–231 tokens and the audio needs 238–360 more. Several need 550+ tokens total, which does not fit a 400-token window in one request.

Generated at 400 anyway, 4 of 10 lost their endings (WER up to 0.507). The rule that predicts it exactly is gen_budget >= chars * 0.85; every clip that failed it truncated, every clip that met it was clean. At 640 all ten fit and only the metric artefact below remains.

Production is not affected by this — tts_server.py splits long text into parts sized by that same rule, so the live agent never issues an over-long request. It is the standalone harness that had to be widened.

Quality

Transcribed back with Sarvam ASR (saaras:v4, auto language, 28 s windows). Raw WER on this content is not a quality measure: the references carry Devanagari proper nouns and the ASR romanises everything, so a word-perfect rendition still scores an error on every Indic token —

REF  नरसिंहा स्वामी took his avatar in the कृता युगा. Lord श्री हरि blessed…
ASR  Narasimha Swami took his avatar in the Krita Yuga. Lord Sri Hari blessed…

That is five "errors" in nine words with nothing wrong. Raw WER therefore tracked Devanagari density rather than anything audible. Folding both sides to one script (ISO-15919, diacritics and vowel length dropped) gives the real figure:

	raw	folded
median WER	0.124	0.066
mean WER	0.157	0.059
coverage	—	1.00 on all 10

clip	Devanagari chars	raw WER	folded WER	dBFS
rag_00	77	0.188	0.087	−12.17
rag_01	23	0.066	0.000	−11.88
rag_02	53	0.127	0.095	−13.52
rag_03	38	0.111	0.037	−12.70
rag_04	49	0.208	0.104	−12.13
rag_05	45	0.111	0.000	−24.87
rag_06	13	0.031	0.015	−26.22
rag_07	48	0.121	0.045	−15.33
rag_08	110	0.415	0.113	−13.07
rag_09	78	0.194	0.090	−13.36

Known issues in these clips

• Level is not uniform: −15.53 ± 5.10 dBFS, a 14.3 dB range. rag_05 and rag_06 come out 12–14 dB below the others. This is content-dependent, not caused by the optimisations — the same two clips were equally quiet in the 400-token run (−24.64, −26.24). Worth a listen and possibly a normalisation pass before any of this reaches users.
• Folded WER 0.066 is higher than the 0.013 the neutral benchmark passages get. Some of that is genuinely harder material (dense proper nouns), and some is the folding being approximate — it flattens consonant distinctions fairly aggressively, so it under- and over-counts in both directions.
• Speaker similarity was not measured. Nothing here establishes that the voice still matches the reference; that needs a listen or a WavLM check.

Contents

• rag_00.wav … rag_09.wav — 48 kHz mono, 16-bit, 28.5–43.5 s
• manifest.json — per clip: text, prompt tokens, generation budget, seed, duration, dBFS, ASR transcript and raw WER