A faster English path is not a universal upgrade.
A useful English result
The 597-character English fixture took 154.75 seconds with sequential Transformers and 122.07 seconds with vLLM batch four: 21.1% less wall time in this single trial. Sequential vLLM was slower at 174.38 seconds. The improvement came from this batched serving configuration, not simply switching engines.
Why the other languages kept their runtime
Luganda took 73.82 seconds with Transformers versus 135.24 seconds with vLLM batch four. Audio length also changed from 19.71 to 31.83 seconds, so the outputs were not interchangeable. The long Swahili baseline failed on its first chunk at roughly 105 seconds; vLLM completed with a longer overall deadline. Completion under different deadlines is not evidence of a speedup.
Decision after listening
Following the comparison, the product owner listened to and accepted the longer English sample. The pilot adopted a conservative 600-character English path, while Luganda and Swahili retained their existing runtime and 300-character limit. This is acceptance of a sample, not a general unchanged-quality finding. Short English probes did not improve with vLLM.
The best long-English configuration still allocated roughly USD 39.65 of busy host compute per million input characters. A subsidised API rate is separate from this measured cost.
Results, in context.
| Language | Runtime / batch | Characters | Generation (s) | Audio (s) |
|---|---|---|---|---|
| EN | transformers-b1 | 597 | 154.75 | 40.28 |
| EN | vllm-b1 | 597 | 174.38 | 40.79 |
| EN | vllm-b2 | 597 | 152.28 | 40.45 |
| EN | vllm-b4 | 597 | 122.07 | 40.87 |
| LG | transformers-b1 | 349 | 73.82 | 19.71 |
| LG | vllm-b1 | 349 | 143.56 | 31.40 |
| LG | vllm-b2 | 349 | 135.02 | 31.83 |
| LG | vllm-b4 | 349 | 135.24 | 31.83 |
| SW | vllm-b1 | 391 | 211.35 | 44.12 |
| SW | vllm-b2 | 391 | 173.44 | 44.20 |
| SW | vllm-b4 | 391 | 174.23 | 44.20 |
Lower generation time is preferable only when content and voice are acceptable. Missing Transformers Swahili row: failed/truncated, not a zero-second result.
How we ran it.
- Same Cape Town T4, pinned Orpheus weights, FP16 and the same voice IDs. Compared the existing Transformers runtime with isolated vLLM 0.10.2 using PyTorch 2.8.
- Long text was split at sentence boundaries into chunks of at most 240 characters and joined in order. English used three chunks; Luganda and Swahili used two. vLLM batches of one, two and four were compared. No time stretching, silence trimming or weight quantization was introduced.
The limits matter.
- Small diagnostic samples, not a production SLA or a population accuracy estimate. Different output lengths affect completion time.
- One long fixture per language/configuration; not a repeated speed estimate. Samples vary in audio duration. No blind listening panel, MOS score or broad language-quality conclusion.
- The Swahili failure remains part of the record and must not be removed from a completion comparison.
- Busy-compute estimates allocate the whole host rate to the measured operation. They exclude idle time, startup, storage, network and taxes; they are not customer API prices or marginal savings. Rates are the recorded October 2026 inputs, not a current quote.
Evidence & sources.
The public extract contains the methods, selected results and caveats in this note. It is a curated summary, not the full raw run. Fingerprints identify the retained source artifacts; gated audio and private operational records are not included.
Download results extractSource artifact fingerprints · SHA-256
78a3aeccd82b029fff82184eb7dbfc65d4c8d5267474920bbf2d1812d61c41c5af6c89c02e774eac799e778f4ab3f749d507715a1ef7cd39760ae4d20d6be153Suggested citation: Embiro Labs Research. “Faster speech, one language at a time.” Published 12 Oct 2026. embiro.ai/research/orpheus-long-form-speech
Report a correction
