Merge pull request #11 from shivenmian/u/shivenmian/local

Integrated local whisper.cpp

.gitignore

@@ -1,3 +1,5 @@
+ggml-*.bin
+
 # Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]

OS/01/conversations/user.json

@@ -1 +1 @@
-[{"role": "user", "type": "message", "content": "There's no fence, but it's like independent units.\n"}]
+[{"role": "user", "type": "message", "content": " Hey, how you doing?\n"}]

OS/01/local_stt/whisper-rust/.gitignore

@@ -0,0 +1,10 @@
+# Generated by Cargo
+# will have compiled files and executables
+debug/
+target/
+
+# These are backup files generated by rustfmt
+**/*.rs.bk
+
+# MSVC Windows builds of rustc generate these, which store debugging information
+*.pdb

OS/01/local_stt/whisper-rust/Cargo.toml

@@ -0,0 +1,14 @@
+[package]
+name = "whisper-rust"
+version = "0.1.0"
+edition = "2021"
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+anyhow = "1.0.79"
+clap = { version = "4.4.18", features = ["derive"] }
+cpal = "0.15.2"
+hound = "3.5.1"
+whisper-rs = "0.10.0"
+whisper-rs-sys = "0.8.0"

OS/01/local_stt/whisper-rust/README.md

@@ -0,0 +1,7 @@
+# Setup
+
+To rebuild the `whisper-rust` executable, do the following:
+
+1. Install [Rust](https://www.rust-lang.org/tools/install), cmake, and Python dependencies `pip install -r requirements.txt`.
+2. Go to **core/stt** and run `cargo build --release`.
+3. Move the `whisper-rust` executable from target/release to this directory.

OS/01/local_stt/whisper-rust/src/main.rs

@@ -0,0 +1,34 @@
+mod transcribe;
+
+use clap::Parser;
+use std::path::PathBuf;
+use transcribe::transcribe;
+
+#[derive(Parser, Debug)]
+#[command(author, version, about, long_about = None)]
+struct Args {
+    /// This is the model for Whisper STT
+    #[arg(short, long, value_parser, required = true)]
+    model_path: PathBuf,
+    
+    /// This is the wav audio file that will be converted from speech to text
+    #[arg(short, long, value_parser, required = true)]
+    file_path: Option<PathBuf>,
+}
+
+fn main() {
+
+    let args = Args::parse();
+
+    let file_path = match args.file_path {
+        Some(fp) => fp,
+        None => panic!("No file path provided")
+    };
+
+    let result = transcribe(&args.model_path, &file_path);
+
+    match result {
+        Ok(transcription) => print!("{}", transcription),
+        Err(e) => panic!("Error: {}", e),
+    }
+}

OS/01/local_stt/whisper-rust/src/transcribe.rs

@@ -0,0 +1,64 @@
+use whisper_rs::{FullParams, SamplingStrategy, WhisperContext, WhisperContextParameters};
+use std::path::PathBuf;
+
+
+/// Transcribes the given audio file using the whisper-rs library.
+///
+/// # Arguments
+/// * `model_path` - Path to Whisper model file
+/// * `file_path` - A string slice that holds the path to the audio file to be transcribed.
+///
+/// # Returns
+///
+/// A Result containing a String with the transcription if successful, or an error message if not.
+pub fn transcribe(model_path: &PathBuf, file_path: &PathBuf) -> Result<String, String> {
+
+    let model_path_str = model_path.to_str().expect("Not valid model path");
+    // Load a context and model
+    let ctx = WhisperContext::new_with_params(
+        model_path_str, // Replace with the actual path to the model
+        WhisperContextParameters::default(),
+    )
+    .map_err(|_| "failed to load model")?;
+
+    // Create a state
+    let mut state = ctx.create_state().map_err(|_| "failed to create state")?;
+
+    // Create a params object
+    // Note that currently the only implemented strategy is Greedy, BeamSearch is a WIP
+    let mut params = FullParams::new(SamplingStrategy::Greedy { best_of: 1 });
+
+    // Edit parameters as needed
+    params.set_n_threads(1); // Set the number of threads to use
+    params.set_translate(true); // Enable translation
+    params.set_language(Some("en")); // Set the language to translate to English
+    // Disable printing to stdout
+    params.set_print_special(false);
+    params.set_print_progress(false);
+    params.set_print_realtime(false);
+    params.set_print_timestamps(false);
+
+    // Load the audio file
+    let audio_data = std::fs::read(file_path)
+        .map_err(|e| format!("failed to read audio file: {}", e))?
+        .chunks_exact(2)
+        .map(|chunk| i16::from_ne_bytes([chunk[0], chunk[1]]))
+        .collect::<Vec<i16>>();
+
+    // Convert the audio data to the required format (16KHz mono i16 samples)
+    let audio_data = whisper_rs::convert_integer_to_float_audio(&audio_data);
+
+    // Run the model
+    state.full(params, &audio_data[..]).map_err(|_| "failed to run model")?;
+
+    // Fetch the results
+    let num_segments = state.full_n_segments().map_err(|_| "failed to get number of segments")?;
+    let mut transcription = String::new();
+    for i in 0..num_segments {
+        let segment = state.full_get_segment_text(i).map_err(|_| "failed to get segment")?;
+        transcription.push_str(&segment);
+        transcription.push('\n');
+    }
+
+    Ok(transcription)
+}

OS/01/start.sh

@@ -1,7 +1,9 @@
 ### SETTINGS
 
 # If ALL_LOCAL is False, we'll use OpenAI's services
+# If setting ALL_LOCAL to true, set the path to the WHISPER local model
 export ALL_LOCAL=False
+# export WHISPER_MODEL_PATH=...
 # export OPENAI_API_KEY=sk-...
 
 # If SERVER_START, this is where we'll serve the server.

OS/01/stt.py

@@ -44,6 +44,28 @@ def export_audio_to_wav_ffmpeg(audio: bytearray, mime_type: str) -> str:
         os.remove(input_path)
         os.remove(output_path)
 
+def run_command(command):
+    result = subprocess.run(command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    return result.stdout, result.stderr
+
+def get_transcription_file(wav_file_path: str):
+    model_path = os.getenv("WHISPER_MODEL_PATH")
+    if not model_path:
+        raise EnvironmentError("WHISPER_MODEL_PATH environment variable is not set.")
+
+    output, error = run_command([
+        os.path.join(os.path.dirname(__file__), 'local_stt', 'whisper-rust', 'whisper-rust'),
+        '--model-path', model_path,
+        '--file-path', wav_file_path
+    ])
+
+    print("Exciting transcription result:", output)
+    return output
+
+def get_transcription_bytes(audio_bytes: bytearray, mime_type):
+    with export_audio_to_wav_ffmpeg(audio_bytes, mime_type) as wav_file_path:
+        return get_transcription_file(wav_file_path)
+
 def stt_bytes(audio_bytes: bytearray, mime_type="audio/wav"):
     with export_audio_to_wav_ffmpeg(audio_bytes, mime_type) as wav_file_path:
         return stt_wav(wav_file_path)
@@ -65,8 +87,15 @@ def stt_wav(wav_file_path: str):
         print("Transcription result:", transcript)
         return transcript
     else:
-        # Local whisper here, given `wav_file_path`
+        temp_dir = tempfile.gettempdir()
-        pass
+        output_path = os.path.join(temp_dir, f"output_stt_{datetime.now().strftime('%Y%m%d%H%M%S%f')}.wav")
+        ffmpeg.input(wav_file_path).output(output_path, acodec='pcm_s16le', ac=1, ar='16k').run()
+        try:
+            transcript = get_transcription_file(output_path)
+            print("Transcription result:", transcript)
+        finally:
+            os.remove(output_path)
+        return transcript
 
 def stt(input_data, mime_type="audio/wav"):
     if isinstance(input_data, str):

README.md

@@ -22,11 +22,12 @@ sudo apt-get install portaudio19-dev libav-tools
 python -m pip install -r requirements.txt
 ```
 
-```bash
+If you want to run local speech-to-text from whisper, download the GGML Whisper model from [Huggingface](https://huggingface.co/ggerganov/whisper.cpp). Then in `OS/01/start.sh`, set `ALL_LOCAL=TRUE` and set `WHISPER_MODEL_PATH` to the path of the model.
-cd OS/01
+
-```
+## Usage
 
 ```bash
+cd OS/01
 bash start.sh
 ```