Skip to content

ModelSelection.md

Latest commit

 

History

History
66 lines (49 loc) · 4.38 KB

ModelSelection.md

File metadata and controls

66 lines (49 loc) · 4.38 KB
Raw

Model Selection

Transcribe architecture has two primary components that require model selection

  • Speech to Text
  • LLM Responses

Speech to Text

Speech to text can run locally, through a recorded-window API, or through a persistent streaming API.

Local Speech to Text

Local Speech to Text requires model selection. By default the base model for English is used. This model is downloaded on the first invocation of the application. There are many more models available, though they vary by size and compute power required.

See the help of transcribe python main.py -h for further details on local transcription models.

  -m {tiny,base,small,medium,large-v1,large-v2,large-v3,large}, --model {tiny,base,small,medium,large-v1,large-v2,large-v3,large}
                        Specify the LLM to use for transcription.
                        By default tiny english model is part of the install.
                        tiny multi-lingual model has to be downloaded from the link   https://drive.google.com/file/d/1M4AFutTmQROaE9xk2jPc5Y4oFRibHhEh/view?usp=drive_link
                        base english model has to be downloaded from the link         https://openaipublic.azureedge.net/main/whisper/models/25a8566e1d0c1e2231d1c762132cd20e0f96a85d16145c3a00adf5d1ac670ead/base.en.pt
                        base multi-lingual model has to be downloaded from the link   https://openaipublic.azureedge.net/main/whisper/models/ed3a0b6b1c0edf879ad9b11b1af5a0e6ab5db9205f891f668f8b0e6c6326e34e/base.pt
                        small english model has to be downloaded from the link        https://openaipublic.azureedge.net/main/whisper/models/f953ad0fd29cacd07d5a9eda5624af0f6bcf2258be67c92b79389873d91e0872/small.en.pt
                        small multi-lingual model has to be downloaded from the link  https://openaipublic.azureedge.net/main/whisper/models/9ecf779972d90ba49c06d968637d720dd632c55bbf19d441fb42bf17a411e794/small.pt

                        The models below require higher computing power:


                        medium english model has to be downloaded from the link       https://openaipublic.azureedge.net/main/whisper/models/d7440d1dc186f76616474e0ff0b3b6b879abc9d1a4926b7adfa41db2d497ab4f/medium.en.pt
                        medium multi-lingual model has to be downloaded from the link https://openaipublic.azureedge.net/main/whisper/models/345ae4da62f9b3d59415adc60127b97c714f32e89e936602e85993674d08dcb1/medium.pt
                        large model has to be downloaded from the link                https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt
                        large-v1 model has to be downloaded from the link             https://openaipublic.azureedge.net/main/whisper/models/e4b87e7e0bf463eb8e6956e646f1e277e901512310def2c24bf0e11bd3c28e9a/large-v1.pt
                        large-v2 model has to be downloaded from the link             https://openaipublic.azureedge.net/main/whisper/models/81f7c96c852ee8fc832187b0132e569d6c3065a3252ed18e56effd0b6a73e524/large-v2.pt
                        large-v3 model has to be downloaded from the link             https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt

Online Speech to Text

The existing Whisper file API is enabled with --api while the whisper engine is selected:

python main.py -stt whisper --api

OpenAI Realtime is a separate streaming backend and does not use --api:

python main.py -stt openai-realtime

It uses gpt-live-transcribe, persistent WebSockets, and independent microphone and speaker-loopback sessions. See OpenAI Realtime transcription. Deepgram remains a recorded-window provider in the current application architecture.

LLM Responses

The quality, cost and speed of responses from the LLM depend on the model chosen. Out of the box Transcribe uses gpt-5.6-luna, the cost-efficient GPT-5.6 model, as specified in parameters.yaml.

    ai_model: gpt-5.6-luna

The model ID is not restricted by Transcribe and can be changed in parameters.yaml or override.yaml. Optional temperature and reasoning_effort values should be left null unless supported by the selected model and provider.

Current OpenAI models are documented at https://developers.openai.com/api/docs/models.