From recording to text. From text to voice.

Compare the wait across devices and runtimes.

Embedding, decision and music workloads →

ModelDevice · runtime

qwen3-asr-rs · Metal · BF16

Model languages: KO · EN · JA +

Input recording · shared English sample

10.2 s · English · Qwen3-TTS 1.7B CustomVoice · Aiden

Estimated processing time

≈ 2.4 s

Processing speed vs playback

4.3×

Estimated-timing demonstration · no live model inference.

Start to see the processing sequence.0.0 s / 2.4 s

A shared, pre-generated example—not a comparison of the selected models’ voices or recognition accuracy.

Timing method and runtime conditions

Published processing ratios are applied to the sample duration. Download and model loading are excluded; short-clip overhead, language and text can change actual time.

M4 · Metal · BF16 · EN/ZH · 3–36s · 3 runs/sample · model already loaded

RTF = 0.230 · 10.2 s × 0.230 ≈ 2.4 s

Transcript reveal is illustrative, not measured word-level timing.

Sample audio credits Timing measurement source

For longer audio

1 min audio

≈ 13.8 s

10 min audio

≈ 138.0 s

60 min audio

≈ 828.0 s

A simple projection using the same processing ratio.

Run it on your own device Explore this device

Which speech model fits your task?

Compare supported languages and setup steps in the guides. Configurations with timing records also let you preview the wait in the demo.

Transcribe recordings and meetings Choose Qwen3-ASR or Whisper From a meeting recording to searchable notes Label meeting speakers with Nemotron Turn writing into spoken audio Run and tune Qwen3-TTS Compare TTS models by language and task