Audio

44 audio aliases · TTS (Kokoro / Qwen3-TTS / Chatterbox / VibeVoice / F5 / IndexTTS / VoxCPM / Dia) + STT (Whisper / Parakeet / SenseVoice / Qwen3-ASR / ForcedAligner).

rapid-mlx serves OpenAI-compatible /v1/audio/transcriptions and /v1/audio/speech endpoints — drop-in for any OpenAI SDK client. 44 aliases ship today across two modalities: text-to-speech (Kokoro, Qwen3-TTS, Chatterbox, VibeVoice, F5, IndexTTS, VoxCPM, Dia) and speech-to-text (Whisper, Parakeet, SenseVoice, Qwen3-ASR, ForcedAligner). All aliases route through the same mlx_audio-backed runtime so swapping models is one CLI flag. rapid-mlx pull fetches a voice model's runtime assets as part of the pull, so a machine that pulled while online can synthesise while offline.

family

Audio

aliases

44

lines

13

install

rapid-mlx serve <alias>

OpenAI base URL

http://localhost:8000/v1

Download

Every alias on this page downloads with one command — the pull buttons in the tables below copy it. All 44 aliases on this page are mirrored on the rapid-mlx CDN — with automatic mid-pull fallback to Hugging Face if a mirror file slows down. Weights land in the standard Hugging Face cache, and rapid-mlx serve pulls automatically on first use. Live mirror status →

Usage

The audio aliases route through OpenAI-compatible endpoints — any OpenAI SDK works without per-model config.

# Text-to-speech
curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model": "kokoro", "input": "hello from rapid-mlx"}' \
  --output hello.wav

# Speech-to-text (multipart upload)
curl -X POST http://localhost:8000/v1/audio/transcriptions \
  -H "Content-Type: multipart/form-data" \
  -F file=@hello.wav \
  -F model=whisper-large-v3-turbo

System-wide local dictation

Rapid-MLX Desktop turns the STT aliases on this page into a system-wide Mac dictation tool. Tap Right Command or Right Option in any app, speak, then tap again: Rapid records only during that session, transcribes through the local /v1/audio/transcriptions endpoint, and pastes the result at the cursor. Raw audio is not archived unless the user explicitly enables audio archiving.

The Desktop model picker explains the practical tradeoffs: Whisper Large v3 Turbo is the balanced multilingual default; Qwen3-ASR targets Chinese/English code-switching and vocabulary hints; SenseVoice is the fast Chinese, Cantonese, Japanese and Korean option; Parakeet v2 is English-only, while v3 supports 25 European languages but not Chinese. See the complete local dictation guide or the 中文本地听写指南.

Lines in this family

• Text-to-speech (TTS) · Kokoro · Chatterbox · VibeVoice · VoxCPM · Dia · Qwen3-TTS · F5-TTS · IndexTTS
• Speech-to-text (STT) · Whisper · Parakeet · Qwen3-ASR · SenseVoice · Qwen3-ForcedAligner

Text-to-speech (TTS) · 24 aliases

OpenAI-compat endpoint: POST /v1/audio/speech. Pass model (any alias from the tables below), input (the text to synthesise), and optionally voice (per-family voice id — falls back to the alias's default voice when omitted).

Kokoro · 7 aliases

Lightweight 82M open-weights TTS — the default rapid-mlx TTS pick. Three quants (bf16 / 8-bit / 4-bit) with short- and long-form aliases for each.

Chatterbox · 2 aliases

Multilingual expressive TTS with voice cloning via reference audio. Both fp16 and 4-bit ship.

VibeVoice · 2 aliases

Realtime low-latency 0.5B TTS. Per-language voice caches; en-Grace_woman is the canonical English-female default.

VoxCPM · 1 alias

Chinese/English high-quality TTS, CPM-based. Single alias.

Dia · 1 alias

Multi-speaker dialogue TTS in 4-bit. Best-effort — covered by the probe but mlx_audio support is experimental.

Qwen3-TTS · 8 aliases

Alibaba's TTS line: Base clones a voice from a reference clip plus its transcript; CustomVoice modulates fixed speakers with an instructions string; VoiceDesign builds a voice from a natural-language description alone — and voice_seed brings the exact same narrator back on the next call.

F5-TTS · 1 alias

Pure-MLX expressive voice cloning (no torch) — closes the Chinese expressive-cloning gap that Qwen3-TTS reads flat on.

IndexTTS · 2 aliases

Clones a voice from the reference clip alone — no transcript needed, the least fiddly cloning path.

Speech-to-text (STT) · 20 aliases

OpenAI-compat endpoint: POST /v1/audio/transcriptions. Send multipart/form-data with file (the audio bytes) and model (any alias from the tables below). Optional language hint accepted by Whisper aliases.

Whisper · 8 aliases

OpenAI's multilingual STT family — tiny → large-v3 + large-v3-turbo. whisper-1 is wired as an OpenAI-compat placeholder pointing at large-v3.

Parakeet · 5 aliases

NVIDIA Parakeet TDT 0.6B — v2 is English-only; v3 supports 25 European languages with automatic language detection. Both checkpoints ship.

Qwen3-ASR · 3 aliases

Alibaba's Qwen3-ASR — an audio LLM rather than an encoder-decoder ASR model. Keeps Chinese/English code-switching straight without a decoding hint and takes hotwords as a system_prompt. Bare qwen3-asr resolves to the 1.7B; a 0.6B ships too.

SenseVoice · 2 aliases

FunAudioLLM SenseVoice Small (~234M, non-autoregressive CTC) — the strongest pick for Chinese, Cantonese, Japanese and Korean, and it emits per-segment emotion and audio-event tags alongside the transcript.

Qwen3-ForcedAligner · 2 aliases

Give it audio plus the transcript you already have and it returns per-character start/end times. It never guesses at words, so it cannot mis-hear them — exactly what karaoke captions and beat-synced editing need.

Notes & caveats

• TTS aliases default to a sensible per-family voice but accept any voice supported by the underlying model — pass voice in the JSON body of /v1/audio/speech.
• STT aliases accept both multipart/form-data audio uploads and the OpenAI-style file field on /v1/audio/transcriptions.
• whisper-1 is wired as an OpenAI-compat placeholder — it resolves to whisper-large-v3 so legacy OpenAI client code works unchanged.
• Dia 1.6B is best-effort — covered by the audio probe but mlx_audio support is experimental upstream; report bugs upstream if synth fails.

External reading

• Kokoro 82M on HF
• Whisper large-v3 (MLX) on HF
• Parakeet TDT 0.6B v2 (MLX) on HF
• Parakeet TDT 0.6B v3 upstream model card

Where next