have to patch the npm package to make it work

[salir-admin]

t a/dist/tts/transcribe.js b/dist/tts/transcribe.js
index fd0ed51c1c0eb61f980559bf8e4fa73945fd148c..8e8db15dd6107fc1381a26e1b1f1aa94e0a44802 100644
--- a/dist/tts/transcribe.js
+++ b/dist/tts/transcribe.js
@@ -11,7 +11,6 @@
  * runs locally via ONNX, no cloud round-trip. The pipeline instance is
  * cached per process so repeated calls reuse the loaded model.
  */
-import { pipeline } from '@huggingface/transformers';
 import { spawnSync } from 'node:child_process';
 export const DEFAULT_WHISPER_MODEL = 'Xenova/whisper-base.en';
 /** Whisper's required audio sample rate. */
@@ -22,6 +21,10 @@ let cached = null;
 async function getTranscriber(model) {
     if (cached?.model === model)
         return cached.transcriber;
+    // Lazy import (patched for bun's isolated linker): a static top-level import made every
+    // CLI invocation — even `--help` — load transformers/onnxruntime, whose nested deps do
+    // not resolve under isolated linking. Whisper is only needed when transcription runs.
+    const { pipeline } = await import('@huggingface/transformers');
     const transcriber = (await pipeline('automatic-speech-recognition', model));
     cached = { model, transcriber };
     return transcriber;
diff --git a/dist/tts/engine.js b/dist/tts/engine.js
index 6c0901d..f2af935 100644
--- a/dist/tts/engine.js
+++ b/dist/tts/engine.js
@@ -153,10 +153,11 @@ export function parseWavHeader(wav) {
  * Convert arbitrary audio (MP3, OGG, PCM, etc.) to Argo's WAV format
  * (mono, Float32, 24kHz) using ffmpeg.
  */
-export function convertToWav(audioBuffer) {
+export function convertToWav(audioBuffer, speed = 1) {
     const { execFileSync } = childProcess;
     const result = execFileSync('ffmpeg', [
         '-i', 'pipe:0',
+        ...(speed === 1 ? [] : ['-filter:a', `atempo=${speed}`]),
         '-f', 'wav',
         '-acodec', 'pcm_f32le',
         '-ac', '1',
diff --git a/dist/tts/engine.d.ts b/dist/tts/engine.d.ts
index 21e6fbf..9b31bd7 100644
--- a/dist/tts/engine.d.ts
+++ b/dist/tts/engine.d.ts
@@ -52,7 +52,7 @@ export declare function parseWavHeader(wav: Buffer): WavInfo;
  * Convert arbitrary audio (MP3, OGG, PCM, etc.) to Argo's WAV format
  * (mono, Float32, 24kHz) using ffmpeg.
  */
-export declare function convertToWav(audioBuffer: Buffer): Buffer;
+export declare function convertToWav(audioBuffer: Buffer, speed?: number): Buffer;
 /**
  * Creates a mock TTS engine that produces silent WAV buffers of the given
  * duration and records all calls for test assertions.
diff --git a/dist/tts/engines/elevenlabs.js b/dist/tts/engines/elevenlabs.js
index 4d194d1..3cc2860 100644
--- a/dist/tts/engines/elevenlabs.js
+++ b/dist/tts/engines/elevenlabs.js
@@ -49,7 +49,7 @@ export class ElevenLabsEngine {
         const mp3Buffer = Buffer.concat(chunks);
         // Convert MP3 to Argo WAV format
         const { convertToWav } = await import('../engine.js');
-        return convertToWav(mp3Buffer);
+        return convertToWav(mp3Buffer, options.speed ?? 1.0);
     }
 }
 //# sourceMappingURL=elevenlabs.js.map

[shreyaskarnik]

Thanks for filing this with the actual diff — two separate things in that patch, and they land differently.

1. The transcribe.js lazy import — already fixed, just not released

You're right, and this was a real blocker. A top-level import { pipeline } from '@huggingface/transformers' in src/tts/transcribe.ts is reachable from src/cli.ts, so it forced every command (validate, export, --help) to load transformers + onnxruntime — which is both wasteful and, as you found, unresolvable under bun's isolated linker.

#36 (in review now) moves that import inside getTranscriber() and routes it through a shared importOptional() helper, same shape as your patch. Transcription is opt-in (tts.transcribe), so nothing else pays for it anymore. You should be able to drop that half of your patch on the next release.

2. The ElevenLabs speed drop — a real bug, thank you for catching it

Confirmed on main. TTSEngineOptions.speed is honored by kokoro, openai, sarvam, mlx-audio and transformers, but ElevenLabs and Gemini both call convertToWav(buffer) and discard the option. No test covers it, which is how it slipped. Your atempo approach is the right generic fix and covers both engines at once.

One thing to watch in your local patch in the meantime: a single atempo instance is clamped to 0.5–2.0. speed: 3 makes ffmpeg exit non-zero, and since convertToWav uses execFileSync that surfaces as a hard throw mid-TTS rather than a graceful degrade. When I land this I'll chain instances for out-of-range values (atempo=2.0,atempo=1.5) and add coverage.

Good news on caching: the clip cache key already includes speed, so cached clips won't go stale once the fix ships — the value simply never reached ffmpeg.

Tracking #2 as a fix for the next release. Leaving this open until it lands.

[shreyaskarnik]

The ElevenLabs speed half is fixed and released in v0.38.2 (npm i @argo-video/cli@0.38.2) — #39.

Your atempo approach, with the range handling added: buildAtempoChain() splits anything outside atempo's per-instance 0.5–2.0 window across several instances whose product is the requested speed, so speed: 3 becomes atempo=2,atempo=1.5 rather than making ffmpeg exit non-zero. Gemini had the identical bug and is fixed in the same change.

Two notes on the local patch you can now drop:

• Invalid values (0, negatives, NaN, Infinity) now throw with a clear message. Worth knowing why: the chain divides by each stage factor until the remainder lands in range, so those inputs never converge — unguarded they hang TTS in a synchronous loop with no output and no error.
• Extreme-but-valid values clamp to 0.25–4.0 with a warning instead of throwing.
• No need to clear .argo/<demo>/clips — the cache key already included speed.

The transcribe.js half is not in this release. #36 is still in review, so if you're on bun you'll need to keep that part of your patch for now. Leaving this issue open until #36 lands; I'll follow up here when it does.

Thanks again — the diff made both of these quick to confirm.