Speed Control

Pass speed to generateSpeech or generateConversation to time-stretch the final audio without changing pitch. 1 is unchanged, values below 1 are slower, values above 1 are faster.

Quick Start

import { generateSpeech } from "@speech-sdk/core"

const result = await generateSpeech({
  model: "openai/gpt-4o-mini-tts",
  text: "Hello!",
  voice: "alloy",
  speed: 1.25,
})

result.audio.mediaType // "audio/mpeg"

Range

speed must be a finite number between 0.75 and 1.5. Passing a value outside that range throws RangeError. 1 (or omitting the parameter) is a no-op.

How It Works

The SDK uses a WSOLA-based time-stretching step on mono PCM audio. It decodes the provider's audio, stretches in the time domain, and re-encodes to your requested output format.

• Direct provider path — time-stretch happens locally in the SDK. To avoid a wasted decode/re-encode round-trip, when speed is set the SDK requests a decodable wire format (PCM/WAV) from the provider and applies the final output format conversion as part of the stretch step.
• Gateway path — speed is forwarded in the wire payload to api.speechbase.ai, which applies the stretch server-side. The gateway invariant (one request, one billed call) is preserved.

timestamps and audioDurationMs on the result are scaled by 1 / speed, so word alignment and reported duration match the actual stretched audio.

Output Format

When speed is set without an explicit output, the stretched audio is encoded as mp3 — matching what most providers return natively. Set output explicitly to override:

const result = await generateSpeech({
  model: "openai/gpt-4o-mini-tts",
  text: "Hello!",
  voice: "alloy",
  speed: 0.85,
  output: { format: "wav" },
})

result.audio.mediaType // "audio/wav"

Conversations

generateConversation accepts speed at the top level and per-turn:

import { generateConversation } from "@speech-sdk/core"

const result = await generateConversation({
  model: "elevenlabs/eleven_v3",
  speed: 1.1, // applies to every turn that doesn't set its own
  turns: [
    { voice: "voice-a", text: "Welcome to the show." },
    { voice: "voice-b", text: "Glad to be here.", speed: 0.9 },
    { voice: "voice-a", text: "Let's get started." },
  ],
})

A per-turn speed forces the stitch path on direct providers, since a single multi-speaker request can't carry per-turn stretch settings. On the gateway path, both top-level and per-turn speed are forwarded as-is.

Notes

• Mono only. Stereo or multi-channel input is not stretched; pass mono audio (which is what every supported provider returns).
• Stretching happens in the time domain, so pitch is preserved. Extreme values toward the edges of the supported range may introduce mild artifacts; stay close to 1 for the cleanest result.
• The same stretch primitive is exported from @speech-sdk/core/plugins as timeStretch if you want to apply it outside the SDK pipeline.