Aller au contenu

Alibaba Qwen3 TTS Fast Alibaba (Qwen) · text-to-speech (TTS) API

Headroom

64 ms median and 97 ms P95 time to first audio on Coval (30 days to October 22, 2026): Headroom speed tier.

Benchmark summary: Headroom speed tier on Coval's P95 (headroom to spend elsewhere in the turn), a tie within 10 ms with Gradium TTS and Inworld Realtime TTS-2 Flash on slow turns. On the Speko speed and naturalness frontier: no other model is both faster and rated higher.

Alibaba Qwen3 TTS Fast is an open-weights text-to-speech (TTS) API from Alibaba (Qwen), served by Nari Labs, measured by two independent benchmarks: Coval and Speko.

Speed tier

Headroom

P95 at or under 101 ms. Headroom to spend elsewhere in the turn.

Median time to first audioCoval, 30 days
64 ms
P95 time to first audio (slow turns)Coval, 30 days
97 ms
Naturalness EloSpeko
1536
Listener Elo (95% interval)Artificial Analysis
not rated
Numbers and dates (0 to 1)Speko
0.3
List price per 1M charactersall sources
$10

Data read October 22, 2026. No combined score: methodology

Gradium TTS vs Alibaba Qwen3 TTS Fast Fastest TTS APIs, by speed tier

Speed

On Coval's 30-day window ending October 22, 2026 (3,374 runs from US East), Alibaba Qwen3 TTS Fast starts speaking after 63.9 ms on a typical turn (median) and 97.3 ms on a slow turn (P95, the slowest one in twenty). That P95 is 2nd lowest of the 32 models Coval measures, which puts it in the Headroom tier: headroom to spend elsewhere in the turn. On the median it is a tie (gap under 10 ms) with Gradium TTS and Inworld Realtime TTS-2 Flash. Speko's board plots it at 37 ms time to first audio from US East; its model page reports a 64 ms sweep.

Voice quality

Speko rates its naturalness at 1,536 Elo from blind A/B votes on phone-agent lines (19th highest of the 43 models Speko rates).

Numbers and dates

Speko's pronunciation test on numbers, dates and currency scores it 0.3 on a 0-to-1 scale (15th highest of the 30 models Speko scores). This is the accuracy measure used here; Coval's word error rate separates models too little to compare them.

Price

Speko lists about $10. List prices, as each source prints them.

Provider and host

The benchmarks measure this model through Nari Labs' API, so the latency above is Nari Labs'. Coval's model registry names Alibaba (Qwen) as the creator and Nari Labs as the host.

Benchmark measurements

Hosting

Served by
Nari Labs

Latency and accuracy — Coval (30-day window)

Mean TTFA
69.3 ms
P50 / median time-to-first-audio (TTFA)
63.9 ms
P90 TTFA
82.5 ms
P95 TTFA
97.3 ms
Latency consistency (standard deviation)
19.3 ms
Leading silence
9.8 ms
Time to first byte (network arrival)
59.5 ms
Accuracy: word error rate (WER)
2%
WER basis
pooled
Latency runs
3,374 runs
Accuracy runs
3,374 runs
Window
30d
Coval snapshot
2026-10-22T14:00:19.256741Z
Coval model ID
qwen3-tts-fast
Coval model page
benchmarks.coval.ai
Tagged multilingual
✓
Feature tags
multilingual

Speko

P50 TTFA, US East
64 ms
Naturalness (Elo)
1,536
Pronunciation robustness (0 to 1)
0.3
Voice drift (lower is steadier)
0
Deterministic output
✓
Price per 1M characters
10 USD
Measured on
2026-09-12
Speko model
Qwen3-TTS Fast
Speko model page
benchmarks.speko.ai
Time to first audio, board chart
37 ms

Price (all sources)

Lowest list price per 1M characters
10 USD
Highest list price per 1M characters
10 USD

Frontiers

On the Speko speed and naturalness frontier
✓

Not collected

Similar TTS models