Step-Audio-EditX
tts · reviewed by Stan K. · Source: data snapshot · data as of 2026-11-12
The odd one out: edits existing audio by text instruction — fix a word, change the emotion — plus zero-shot TTS, if your GPU has 12 GB+ to spare.
#12 of 16 in Tts
ToolScore breakdown — why 55/100
10 criteria × 10 points. Low pillars are visible on purpose — that's the point of a score you can audit. Methodology →
What it does well
- ▸Edits existing audio by text instruction — fix a word, change the emotion: a capability nothing else in this table has
- ▸Zero-shot TTS included alongside the editing
- ▸Hardware needs fully published — the only local model in our table where the watts bill is exactly computable
- ▸Apache 2.0: commercial use, no NC trap
Where it costs you
- ▸12 GB+ GPU to start (~6–8 GB AWQ): the steepest published hardware bar in the local table
- ▸3B params — the heaviest model here; the value-per-VRAM math punishes it accordingly
- ▸No latency number and no arena row — speed and quality unverifiable
Hidden traps — the stuff the pricing page doesn't say
~6–8 GB with AWQ quantization, 12 GB+ GPU for the comfortable path. The electricity is ≈$0.03/hour-cheap once you're running — but the model's cost sits in the GPU you must already own, which is exactly what our value-per-VRAM metric measures.
Know the limits before you pay: read the traps above, then decide with eyes open.
Affiliate link where available — never affects the score.
- internal ToolScore TTS hub — local table: Apache 2.0, 3B params, ~6–8 GB (AWQ) · 12 GB+ GPU (licenses from LICENSE files and model cards) verified
Data as of 2026-11-12. Prices and limits change monthly — report anything stale. How we verify: methodology.
Closest alternatives
tts-1 verifies at ≈$0.011/min — but gpt-4o-mini-tts bills $12/1M audio output tokens with no publish
Apache 2.0 two-speaker dialogue in a single pass — [S1]/[S2] interruptions and reactions without sti
≈$0.011/min on the API with Elo 1274 quality — and the local weights are CC-BY-NC, so commercial use
Step-Audio-EditX review
In-depth analysis · last updated 2026-11-12
The editor in a table of speakers
Step-Audio-EditX is the odd one out, and proudly: every other model here generates speech from text, while EditX edits audio that already exists — fix a word, change the emotion — by text instruction, with zero-shot TTS thrown in. In a production pipeline where re-recording a 40-minute audiobook chapter because of one mispronounced name, editing-by-instruction is not a convenience feature; it is a different tool category.
The economics — published, and expensive anyway
Credit where due: this is the only local model in our table whose hardware needs are fully published — ~6–8 GB (AWQ), 12 GB+ GPU — so the watts math is exact rather than guessed. It's also the math that sinks it: 3B params behind a 12 GB+ entry ticket is the heaviest bill in the local table, and the hub sorts locals by value per VRAM. Electricity at ≈$0.03/hour is cheap once running; the GPU you must own to start is the price.
Real limits
No quotas — self-hosted, no meters, Apache 2.0 with no trap. The limits are physical: 12 GB+ of VRAM, AWQ quantization on the practical path, no latency figure published, no arena row. The feature set is two-pronged (edit + zero-shot) but niche-shaped — "the odd one out" is the hub's own description, and it is the accurate one.
How the pillars land
Price/Value 1.5: the value-per-VRAM floor of the healthy locals — a 12 GB+ GPU for a specialist tool. Quotas 6: no meter exists, but practical batch volume on one big GPU is narrower than a CPU fleet. Latency the neutral 5 — no latency fact in our snapshot. Training use & voice cloning 6: local inference, nothing leaves the box; no cloning claims on record. API & self-host 10: Apache 2.0, no trap. Feature set 6.5: audio-editing-by-instruction plus zero-shot TTS — a unique pair, scored for its narrowness. Track record the neutral 5. Ease 3: AWQ plus a 12 GB+ GPU is a real setup bar — the hardest published requirements in the table. Pricing transparency 8: the only fully published hardware sheet among the locals, docked because nothing else (latency, quality reference) is published. Support & docs the neutral 5.
Verdict — NICHE, 55/100
If you re-record audio for a living, the editing capability justifies the GPU and nothing else in this table does the job. As a general TTS engine it's the wrong purchase: too heavy, too specialist. Buy it for the edit button, keep Kokoro for the bulk — and be honest that you're feeding a 3B model to fix single words.