cascade.yaml

# Cascade provider catalog: speech-to-text -> LLM -> text-to-speech
# API keys belong in .env or real environment variables.
#
# This file defines available providers. Gradio setup chooses which providers
# to use; CLI/headless mode may still use the provider keys below as defaults.

asr:
  provider: parakeet_mlx_progressive

  cloud_providers:
    whisper_openai:
      module: whisper_openai
      class: WhisperOpenAIASR
      streaming: false
      requires: [OPENAI_API_KEY]
      hardware: null
      description: "OpenAI Whisper API - high quality cloud batch ASR"
      model: whisper-1
      cost_per_second: 0.0001

    deepgram:
      module: deepgram
      class: DeepgramASR
      streaming: true
      requires: [DEEPGRAM_API_KEY]
      hardware: null
      description: "Deepgram Nova - real-time cloud streaming ASR"
      model: nova-2
      language: en

    openai_realtime_asr:
      module: openai_realtime_asr
      class: OpenAIRealtimeASR
      streaming: true
      requires: [OPENAI_API_KEY]
      hardware: null
      description: "OpenAI Realtime - GPT-4o streaming transcription via WebSocket"
      model: gpt-4o-transcribe
      language: en
      prompt: ""
      use_server_vad: true

  local_providers:
    parakeet_mlx_progressive:
      module: parakeet_mlx_progressive
      class: ParakeetMLXProgressiveASR
      streaming: true
      requires: []
      hardware: apple_silicon
      import_check: mlx_audio
      install_extra: cascade_parakeet_progressive
      description: "Parakeet MLX progressive - live transcription with sentence-aware sliding window"
      model: mlx-community/parakeet-tdt-0.6b-v3
      precision: float16
      max_window_size: 15.0
      sentence_buffer: 2.0

    voxtral_mlx:
      module: voxtral_mlx
      class: VoxtralMLXASR
      streaming: true
      requires: []
      hardware: apple_silicon
      import_check: mlx_audio
      install_extra: cascade_voxtral_mlx
      description: "Voxtral Mini 4B Realtime - multilingual local ASR via MLX"
      model: mlx-community/Voxtral-Mini-4B-Realtime-6bit
      language: en
      max_tokens: 256
      transcription_delay_ms: 480

    parakeet_nemo_progressive:
      module: parakeet_nemo_progressive
      class: ParakeetNeMoProgressiveASR
      streaming: true
      requires: []
      hardware: cuda
      import_check: nemo.collections
      install_extra: cascade_nemotron
      description: "Parakeet TDT via NeMo - progressive ASR with sliding window (CUDA)"
      model: nvidia/parakeet-tdt-0.6b-v2
      max_window_size: 15.0
      sentence_buffer: 2.0

    nemotron:
      module: nemotron
      class: NemotronASR
      streaming: true
      requires: []
      hardware: cuda
      import_check: nemo.collections
      install_extra: cascade_nemotron
      description: "NVIDIA Nemotron - cache-aware streaming ASR"
      model: nvidia/nemotron-speech-streaming-en-0.6b
      chunk_size_ms: 560


llm:
  provider: gemini-2.5-flash-lite
  temperature: 1.0

  cloud_providers:
    gpt-5.2-chat:
      module: openai
      class: OpenAILLM
      requires: [OPENAI_API_KEY]
      description: "OpenAI GPT 5.2 Chat"
      model: gpt-5.2-chat-latest
      input_cost_per_1m: 1.75
      output_cost_per_1m: 14.00

    gpt-4o-mini:
      module: openai
      class: OpenAILLM
      requires: [OPENAI_API_KEY]
      description: "OpenAI GPT models"
      model: gpt-4o-mini
      input_cost_per_1m: 0.15
      output_cost_per_1m: 0.60

    gemini-2.5-flash-lite:
      module: gemini
      class: GeminiLLM
      requires: [GEMINI_API_KEY]
      description: "Google Gemini models"
      model: gemini-2.5-flash-lite
      input_cost_per_1m: 0.075
      output_cost_per_1m: 0.30

    gemini-3.1-flash-lite:
      module: gemini
      class: GeminiLLM
      requires: [GEMINI_API_KEY]
      description: "Google Gemini 3.1 Flash Lite - fastest, most cost-efficient Gemini 3 model"
      model: gemini-3.1-flash-lite-preview
      input_cost_per_1m: 0.25
      output_cost_per_1m: 1.50

  local_providers:
    local_openai_compatible:
      module: openai
      class: OpenAILLM
      requires: []
      description: "Local OpenAI-compatible server (LM Studio, Ollama, vLLM, llama.cpp)"
      base_url: ""
      api_key: ""
      model: ""
      input_cost_per_1m: 0.0
      output_cost_per_1m: 0.0


tts:
  provider: kokoro
  trim_silence: true

  cloud_providers:
    tts_openai:
      module: openai
      class: OpenAITTS
      requires: [OPENAI_API_KEY]
      description: "OpenAI TTS API"
      voice: alloy
      model: tts-1
      cost_per_1m_chars: 15.0

    elevenlabs:
      module: elevenlabs
      class: ElevenLabsTTS
      requires: [ELEVENLABS_API_KEY]
      description: "ElevenLabs TTS API"
      voice_id: pNInz6obpgDQGcFmaJgB
      model: eleven_flash_v2_5
      cost_per_1m_chars: 30.0

    gradium:
      module: gradium
      class: GradiumTTS
      requires: [GRADIUM_API_KEY]
      import_check: gradium
      install_extra: cascade_gradium
      description: "Gradium TTS - low-latency streaming voice AI"
      voice_id: YTpq7expH9539ERJ
      model: default
      cost_per_1m_chars: 0.0

  local_providers:
    local_openai_compatible_tts:
      module: openai
      class: OpenAITTS
      requires: []
      description: "Local OpenAI-compatible TTS server"
      base_url: ""
      api_key: ""
      voice: ""
      model: ""
      response_format: pcm
      cost_per_1m_chars: 0.0

    kokoro:
      module: kokoro
      class: KokoroTTS
      requires: []
      hardware: null
      import_check: kokoro
      install_extra: cascade_kokoro
      description: "Kokoro local TTS - works on any PyTorch platform"
      voice: am_adam


transcript_analysis:
  gliner_model: urchade/gliner_small-v2.1