Spum Voice
Measured speech generation. Upstream baselines. Audio you can compare.
What this run establishesMaya runs at essentially the same speed with Spum’s profile as with matched upstream settings. Spum currently integrates the upstream voice engine; it does not replace Maya’s inference kernels. The CPU comparison evaluates original Kokoro, Spum’s native INT8 Kokoro deployment and the classical eSpeak NG reference.
H200 DGX · GPU 0
| Configuration | Clips | Median RTF ↓ | Median generation | First PCM chunk | English ASR WER ↓ |
|---|
| Maya · default | 10 | 0.135 | 0.843 s | 14.6 ms | 0.0% |
|---|
| Spum Voice · Maya | 10 | 0.135 | 0.773 s | 14.7 ms | 0.0% |
|---|
| Maya · matched sampling | 10 | 0.135 | 0.758 s | 14.7 ms | 0.0% |
|---|
Apple M5 Max · CPU
| Configuration | Clips | Median RTF ↓ | Median generation | First PCM chunk | English ASR WER ↓ |
|---|
| eSpeak NG reference | 6 | 0.002 | 0.011 s | — | 7.3% |
|---|
| Kokoro upstream FP32 | 6 | 0.088 | 0.679 s | — | 0.0% |
|---|
| Spum Voice Kokoro INT8 | 6 | 0.637 | 4.194 s | — | 0.0% |
|---|
RTF is generation time divided by audio duration; lower is faster. Results are sorted within each hardware group. The Maya run uses one checkpoint and Ritvik voice, five prompts, two repetitions per condition, a fixed seed and an excluded warm-up. “Matched” uses Spum’s sampling settings without the prosody preprocessing. First-chunk timing is the first non-empty PCM buffer, not measured audible onset.
Kokoro uses three English prompts and two repetitions. Original Kokoro is a warm persistent FP32 CPU model. Spum uses INT8 on the same CPU with four threads; its wall time includes process startup and model loading. This comparison does not isolate kernel speed or quantization quality. eSpeak NG is a lightweight classical synthesizer, not a comparable neural voice.
ASR WER uses Whisper base.en and the upstream Whisper English normalizer; lower is better. It is a small-sample transcription proxy, not a human quality score. Hindi naturalness and pronunciation have not been assigned automatic quality ratings. The small Maya timing differences do not establish a speed winner. The CPU configurations differ in precision and process lifetime.
Listen without the labels
Rate naturalness and clarity from 1 to 5 before revealing the configurations. Ratings stay in this browser.
The scanner is ready. The image has been saved. Please review the result before moving the optical stage.
Voice ASpum Voice · MayaVoice BMaya · defaultVoice CMaya · matched sampling The exposure is eighty milliseconds. The signal to noise ratio is seventy one point one, at twelve point five frames per second.
Voice AMaya · defaultVoice BSpum Voice · MayaVoice CMaya · matched sampling The integral from zero to infinity of exponential minus x squared is the square root of pi, divided by two.
Voice AMaya · matched samplingVoice BMaya · defaultVoice CSpum Voice · Maya अगर बल की रेखा केंद्र से गुजरती है, तो उस बल का टॉर्क शून्य क्यों होता है?
Voice ASpum Voice · MayaVoice BMaya · matched samplingVoice CMaya · default बल से वस्तु का आगे बढ़ना शुरू होता है, लेकिन बिना फिसले घूमने के लिए स्थैतिक घर्षण टॉर्क देता है।
Voice ASpum Voice · MayaVoice BMaya · defaultVoice CMaya · matched sampling The scanner is ready. The image has been saved. Please review the result before moving the optical stage.
Voice AKokoro upstream FP32Voice BeSpeak NG referenceVoice CSpum Voice Kokoro INT8 The exposure is eighty milliseconds. The signal to noise ratio is seventy one point one, at twelve point five frames per second.
Voice AKokoro upstream FP32Voice BSpum Voice Kokoro INT8Voice CeSpeak NG reference The integral from zero to infinity of exponential minus x squared is the square root of pi, divided by two.
Voice AeSpeak NG referenceVoice BKokoro upstream FP32Voice CSpum Voice Kokoro INT8 Other open models reviewed
Chatterbox and Qwen3-TTS are candidates for a broader neural comparison. They were not executed in this run, so no performance ranking is assigned.
Reproducibility and attribution
Download timings, parameters, transcripts and result data. Audio files are linked in the report. Upstream models and code retain their licences: Maya, Kokoro, eSpeak NG. Maya’s Mimi component requires CC BY attribution; see the upstream model card.