| 1 | //! Voice capture and transcription core: recorder detection, WAV encoding, |
| 2 | //! ASR selection and dispatch, the send-suffix contract, and the headless |
| 3 | //! record→transcribe cycle. |
| 4 | //! |
| 5 | //! Shared by the TUI's `/voice` commands (`commands::voice`, which keeps the |
| 6 | //! slash commands and the composer-updating capture loop) and the runtime |
| 7 | //! API's `/v1/voice` routes (`runtime_api::voice`). No UI types live here. |
| 8 | //! |
| 9 | //! ## Recording |
| 10 | //! |
| 11 | //! Uses platform-specific command-line tools (sox, rec, arecord) to capture |
| 12 | //! 16kHz mono 16-bit PCM audio. Records until a silence gap is detected or |
| 13 | //! the maximum duration is reached (default 10 s). |
| 14 | |
| 15 | use std::process::{Command, Stdio}; |
| 16 | use std::sync::LazyLock; |
| 17 | use std::time::Duration; |
| 18 | |
| 19 | use regex::Regex; |
| 20 | |
| 21 | use crate::config::Config; |
| 22 | |
| 23 | /// Transcription model requested from the provider's chat-completions API. |
| 24 | const ASR_MODEL: &str = "mimo-v2.5-asr"; |
| 25 | const GROQ_ASR_MODEL: &str = "whisper-large-v3-turbo"; |
| 26 | /// Local whisper binary names to probe (whisper.cpp, faster-whisper, OpenAI whisper). |
| 27 | const LOCAL_WHISPER_BINS: &[&str] = &["whisper", "whisper.cpp", "whisper-cpp", "faster-whisper"]; |
| 28 | /// Model used for the AI-assisted voice-control pipeline. |
| 29 | const VOICE_CONTROL_MODEL: &str = "mimo-v2.5"; |
| 30 | |
| 31 | // --- Recorder detection ---------------------------------------------------- |
| 32 | |
| 33 | /// Platform-specific recorder definitions. |
| 34 | #[derive(Debug, Clone)] |
| 35 | struct Recorder { |
| 36 | cmd: &'static str, |
| 37 | /// CLI arguments for piping raw 16kHz mono S16_LE PCM to stdout. |
| 38 | pipe_args: &'static [&'static str], |
| 39 | } |
| 40 | |
| 41 | fn detect_recorder() -> Option<Recorder> { |
| 42 | // Operator kill-switch: a headless `serve --http` host has no business |
| 43 | // opening a microphone; disabling voice here makes `GET /v1/voice` |
| 44 | // report `available: false` and every dictate call fail closed. |
| 45 | if std::env::var_os("CODEWHALE_DISABLE_VOICE").is_some() { |
| 46 | return None; |
| 47 | } |
| 48 | let candidates: &[Recorder] = if cfg!(target_os = "macos") { |
| 49 | &[ |
| 50 | Recorder { |
| 51 | cmd: "sox", |
| 52 | pipe_args: &["-d", "-r", "16000", "-c", "1", "-b", "16", "-t", "raw", "-"], |
| 53 | }, |
| 54 | Recorder { |
| 55 | cmd: "rec", |
| 56 | pipe_args: &["-r", "16000", "-c", "1", "-b", "16", "-t", "raw", "-"], |
| 57 | }, |
| 58 | ] |
| 59 | } else if cfg!(target_os = "linux") { |
| 60 | &[ |
| 61 | Recorder { |
| 62 | cmd: "arecord", |
| 63 | pipe_args: &["-f", "S16_LE", "-r", "16000", "-c", "1", "-t", "raw"], |
| 64 | }, |
| 65 | Recorder { |
| 66 | cmd: "sox", |
| 67 | pipe_args: &["-d", "-r", "16000", "-c", "1", "-b", "16", "-t", "raw", "-"], |
| 68 | }, |
| 69 | ] |
| 70 | } else if cfg!(target_os = "windows") { |
| 71 | &[Recorder { |
| 72 | cmd: "sox", |
| 73 | pipe_args: &["-d", "-r", "16000", "-c", "1", "-b", "16", "-t", "raw", "-"], |
| 74 | }] |
| 75 | } else { |
| 76 | &[] |
| 77 | }; |
| 78 | |
| 79 | candidates |
| 80 | .iter() |
| 81 | .find(|r| { |
| 82 | Command::new(r.cmd) |
| 83 | .arg("--version") |
| 84 | .stdin(Stdio::null()) |
| 85 | .stdout(Stdio::null()) |
| 86 | .stderr(Stdio::null()) |
| 87 | .spawn() |
| 88 | .is_ok() |
| 89 | }) |
| 90 | .cloned() |
| 91 | } |
| 92 | |
| 93 | /// Check whether voice recording is available on this system. |
| 94 | pub fn is_available() -> bool { |
| 95 | detect_recorder().is_some() |
| 96 | } |
| 97 | |
| 98 | // --- WAV encoding ---------------------------------------------------------- |
| 99 | |
| 100 | /// Encode raw 16kHz mono S16_LE PCM samples as a WAV buffer. |
| 101 | fn encode_wav(samples: &[i16]) -> Vec<u8> { |
| 102 | let data_size = (samples.len() * 2) as u32; |
| 103 | let sample_rate: u32 = 16000; |
| 104 | let mut buf = Vec::with_capacity(44 + data_size as usize); |
| 105 | |
| 106 | // RIFF header |
| 107 | buf.extend_from_slice(b"RIFF"); |
| 108 | buf.extend_from_slice(&(36 + data_size).to_le_bytes()); |
| 109 | buf.extend_from_slice(b"WAVE"); |
| 110 | |
| 111 | // fmt chunk |
| 112 | buf.extend_from_slice(b"fmt "); |
| 113 | buf.extend_from_slice(&16u32.to_le_bytes()); // chunk size |
| 114 | buf.extend_from_slice(&1u16.to_le_bytes()); // PCM |
| 115 | buf.extend_from_slice(&1u16.to_le_bytes()); // mono |
| 116 | buf.extend_from_slice(&sample_rate.to_le_bytes()); |
| 117 | buf.extend_from_slice(&(sample_rate * 2).to_le_bytes()); // byte rate |
| 118 | buf.extend_from_slice(&2u16.to_le_bytes()); // block align |
| 119 | buf.extend_from_slice(&16u16.to_le_bytes()); // bits per sample |
| 120 | |
| 121 | // data chunk |
| 122 | buf.extend_from_slice(b"data"); |
| 123 | buf.extend_from_slice(&data_size.to_le_bytes()); |
| 124 | for &sample in samples { |
| 125 | buf.extend_from_slice(&sample.to_le_bytes()); |
| 126 | } |
| 127 | |
| 128 | buf |
| 129 | } |
| 130 | |
| 131 | // --- Recording ------------------------------------------------------------- |
| 132 | |
| 133 | /// Maximum recording duration in seconds before auto-stopping. |
| 134 | pub const MAX_RECORD_SECS: u64 = 10; |
| 135 | /// Minimum segment duration in seconds to consider as valid speech. |
| 136 | const MIN_SEGMENT_SECS: f64 = 0.3; |
| 137 | |
| 138 | /// Record audio from the default microphone. |
| 139 | /// |
| 140 | /// Returns raw 16kHz mono S16_LE PCM samples. Returns `None` if no recorder |
| 141 | /// is available, the recording failed, or no speech was detected. |
| 142 | pub(crate) fn record_audio() -> Option<(Vec<i16>, Duration)> { |
| 143 | let recorder = detect_recorder()?; |
| 144 | let start = std::time::Instant::now(); |
| 145 | |
| 146 | let mut child = Command::new(recorder.cmd) |
| 147 | .args(recorder.pipe_args) |
| 148 | .stdin(Stdio::null()) |
| 149 | .stdout(Stdio::piped()) |
| 150 | .stderr(Stdio::null()) |
| 151 | .spawn() |
| 152 | .ok()?; |
| 153 | |
| 154 | let stdout = child.stdout.take()?; |
| 155 | let mut reader = std::io::BufReader::new(stdout); |
| 156 | let mut all_samples: Vec<i16> = Vec::with_capacity(16000 * MAX_RECORD_SECS as usize); |
| 157 | |
| 158 | // Read until timeout or silence |
| 159 | let mut buf = [0u8; 320]; // 10ms of 16kHz S16_LE |
| 160 | let max_duration = Duration::from_secs(MAX_RECORD_SECS); |
| 161 | let mut silence_samples = 0u32; |
| 162 | let mut had_speech = false; |
| 163 | let speech_threshold: i16 = 500; // RMS-based speech detection threshold |
| 164 | let silence_duration_samples = 16000u32; // 1 second of silence to stop |
| 165 | |
| 166 | loop { |
| 167 | use std::io::Read; |
| 168 | match reader.read_exact(&mut buf) { |
| 169 | Ok(()) => { |
| 170 | let chunk: Vec<i16> = buf |
| 171 | .as_chunks::<2>() |
| 172 | .0 |
| 173 | .iter() |
| 174 | .copied() |
| 175 | .map(i16::from_le_bytes) |
| 176 | .collect(); |
| 177 | |
| 178 | // Simple RMS-based VAD |
| 179 | let rms = (chunk.iter().map(|&s| (s as f64) * (s as f64)).sum::<f64>() |
| 180 | / chunk.len() as f64) |
| 181 | .sqrt(); |
| 182 | let is_speech = rms > speech_threshold as f64; |
| 183 | |
| 184 | if is_speech { |
| 185 | had_speech = true; |
| 186 | silence_samples = 0; |
| 187 | } else if had_speech { |
| 188 | silence_samples += chunk.len() as u32; |
| 189 | } |
| 190 | |
| 191 | if had_speech { |
| 192 | all_samples.extend_from_slice(&chunk); |
| 193 | } |
| 194 | |
| 195 | if start.elapsed() > max_duration { |
| 196 | let _ = child.kill(); |
| 197 | break; |
| 198 | } |
| 199 | if had_speech && silence_samples >= silence_duration_samples { |
| 200 | let _ = child.kill(); |
| 201 | break; |
| 202 | } |
| 203 | } |
| 204 | Err(e) if e.kind() == std::io::ErrorKind::UnexpectedEof => break, |
| 205 | Err(_) => { |
| 206 | let _ = child.kill(); |
| 207 | break; |
| 208 | } |
| 209 | } |
| 210 | } |
| 211 | |
| 212 | let _ = child.wait(); |
| 213 | let elapsed = start.elapsed(); |
| 214 | |
| 215 | let min_samples = (MIN_SEGMENT_SECS * 16000.0) as usize; |
| 216 | if all_samples.len() < min_samples { |
| 217 | return None; |
| 218 | } |
| 219 | |
| 220 | Some((all_samples, elapsed)) |
| 221 | } |
| 222 | |
| 223 | // --- Auto-send suffix ------------------------------------------------------ |
| 224 | |
| 225 | /// Trailing phrases that mean "submit this" — the human-readable form of |
| 226 | /// `SEND_SUFFIX_RE`; keep in sync with the regex when either changes. |
| 227 | pub const SEND_PHRASES: &[&str] = &["send it", "发送", "發送"]; |
| 228 | |
| 229 | /// Matches an explicit send instruction at the end of transcribed text: |
| 230 | /// "send it" (any spacing/case) or 发送/發送, with trailing punctuation. |
| 231 | static SEND_SUFFIX_RE: LazyLock<Regex> = LazyLock::new(|| { |
| 232 | Regex::new(r"(?i)(?:^|[\s,，.。!！?？]+)(?:send\s*it|发送|發送)[\s.。!！?？]*$").unwrap() |
| 233 | }); |
| 234 | |
| 235 | /// Split a transcript into the message remainder and whether it ended with an |
| 236 | /// explicit send instruction. `"ship the fix, send it"` → `("ship the fix", true)`. |
| 237 | pub(crate) fn split_send_suffix(text: &str) -> (&str, bool) { |
| 238 | match SEND_SUFFIX_RE.find(text) { |
| 239 | Some(found) => (text[..found.start()].trim(), true), |
| 240 | None => (text.trim(), false), |
| 241 | } |
| 242 | } |
| 243 | |
| 244 | // --- Transcription --------------------------------------------------------- |
| 245 | |
| 246 | fn base64_encode(data: &[u8]) -> String { |
| 247 | use base64::Engine; |
| 248 | base64::engine::general_purpose::STANDARD.encode(data) |
| 249 | } |
| 250 | |
| 251 | fn chat_completions_url(base_url: &str) -> String { |
| 252 | format!("{}/chat/completions", base_url.trim_end_matches('/')) |
| 253 | } |
| 254 | |
| 255 | async fn post_chat_completions( |
| 256 | api_key: &str, |
| 257 | base_url: &str, |
| 258 | mut body: serde_json::Value, |
| 259 | openrouter_vendor: Option<&str>, |
| 260 | ) -> Result<serde_json::Value, String> { |
| 261 | crate::client::apply_openrouter_vendor(&mut body, openrouter_vendor); |
| 262 | let _inference = crate::client::acquire_remote_control_inference_participant().await; |
| 263 | let client = crate::tls::reqwest_client(); |
| 264 | let resp = client |
| 265 | .post(chat_completions_url(base_url)) |
| 266 | .header("Content-Type", "application/json") |
| 267 | .header("Authorization", format!("Bearer {api_key}")) |
| 268 | .timeout(Duration::from_secs(30)) |
| 269 | .json(&body) |
| 270 | .send() |
| 271 | .await |
| 272 | .map_err(|e| format!("request failed: {e}"))?; |
| 273 | |
| 274 | if !resp.status().is_success() { |
| 275 | return Err(format!("API returned status {}", resp.status())); |
| 276 | } |
| 277 | |
| 278 | resp.json() |
| 279 | .await |
| 280 | .map_err(|e| format!("failed to parse response: {e}")) |
| 281 | } |
| 282 | |
| 283 | /// Send audio to the provider's API for plain transcription. |
| 284 | /// |
| 285 | /// Uses the chat completions endpoint with `input_audio` content blocks. |
| 286 | pub(crate) async fn transcribe( |
| 287 | api_key: &str, |
| 288 | base_url: &str, |
| 289 | audio_samples: &[i16], |
| 290 | openrouter_vendor: Option<&str>, |
| 291 | ) -> Result<String, String> { |
| 292 | transcribe_with_model( |
| 293 | api_key, |
| 294 | base_url, |
| 295 | audio_samples, |
| 296 | ASR_MODEL, |
| 297 | openrouter_vendor, |
| 298 | ) |
| 299 | .await |
| 300 | } |
| 301 | |
| 302 | async fn transcribe_with_model( |
| 303 | api_key: &str, |
| 304 | base_url: &str, |
| 305 | audio_samples: &[i16], |
| 306 | model: &str, |
| 307 | openrouter_vendor: Option<&str>, |
| 308 | ) -> Result<String, String> { |
| 309 | let wav = encode_wav(audio_samples); |
| 310 | let data_url = format!("data:audio/wav;base64,{}", base64_encode(&wav)); |
| 311 | |
| 312 | let body = serde_json::json!({ |
| 313 | "model": model, |
| 314 | "messages": [ |
| 315 | { |
| 316 | "role": "user", |
| 317 | "content": [ |
| 318 | { |
| 319 | "type": "input_audio", |
| 320 | "input_audio": { |
| 321 | "data": data_url |
| 322 | } |
| 323 | } |
| 324 | ] |
| 325 | } |
| 326 | ], |
| 327 | "asr_options": { |
| 328 | "language": "auto" |
| 329 | } |
| 330 | }); |
| 331 | |
| 332 | let data = post_chat_completions(api_key, base_url, body, openrouter_vendor).await?; |
| 333 | data["choices"][0]["message"]["content"] |
| 334 | .as_str() |
| 335 | .map(|s| s.trim().to_string()) |
| 336 | .ok_or_else(|| "no transcription in response".to_string()) |
| 337 | } |
| 338 | |
| 339 | /// Process audio through the voice-control pipeline: AI-assisted dictation |
| 340 | /// that sees the current composer text, mirroring MiMo Code's |
| 341 | /// `processVoiceControl`. Used when `/voice-control` is enabled. |
| 342 | pub(crate) async fn process_voice_control( |
| 343 | api_key: &str, |
| 344 | base_url: &str, |
| 345 | audio_samples: &[i16], |
| 346 | current_text: &str, |
| 347 | openrouter_vendor: Option<&str>, |
| 348 | ) -> Result<String, String> { |
| 349 | let wav = encode_wav(audio_samples); |
| 350 | let data_url = format!("data:audio/wav;base64,{}", base64_encode(&wav)); |
| 351 | |
| 352 | let user_context = serde_json::json!({ |
| 353 | "current_text": current_text, |
| 354 | "cursor": "end", |
| 355 | }); |
| 356 | |
| 357 | let body = serde_json::json!({ |
| 358 | "model": VOICE_CONTROL_MODEL, |
| 359 | "messages": [ |
| 360 | { |
| 361 | "role": "system", |
| 362 | "content": "You are a voice input assistant. Transcribe the user's speech. Output JSON: {\"text\": \"transcribed text\"}." |
| 363 | }, |
| 364 | { |
| 365 | "role": "user", |
| 366 | "content": [ |
| 367 | { "type": "text", "text": user_context.to_string() }, |
| 368 | { "type": "input_audio", "input_audio": { "data": data_url } } |
| 369 | ] |
| 370 | } |
| 371 | ], |
| 372 | "response_format": { "type": "json_object" } |
| 373 | }); |
| 374 | |
| 375 | let data = post_chat_completions(api_key, base_url, body, openrouter_vendor).await?; |
| 376 | let content = data["choices"][0]["message"]["content"] |
| 377 | .as_str() |
| 378 | .ok_or_else(|| "no response content".to_string())?; |
| 379 | |
| 380 | let parsed: serde_json::Value = serde_json::from_str(content) |
| 381 | .map_err(|e| format!("failed to parse voice control JSON: {e}"))?; |
| 382 | |
| 383 | parsed["text"] |
| 384 | .as_str() |
| 385 | .map(|s| s.to_string()) |
| 386 | .ok_or_else(|| "no text field in voice control response".to_string()) |
| 387 | } |
| 388 | |
| 389 | /// Detect best free ASR for this host — local whisper > Groq free > provider fallback. |
| 390 | /// Works on macOS (brew install whisper-cpp), Windows (whisper.cpp binary), |
| 391 | /// Linux (apt), and HarmonyOS (falls back to cloud). |
| 392 | fn detect_free_asr() -> &'static str { |
| 393 | for bin in LOCAL_WHISPER_BINS { |
| 394 | if Command::new(bin) |
| 395 | .arg("--help") |
| 396 | .stdin(Stdio::null()) |
| 397 | .stdout(Stdio::null()) |
| 398 | .stderr(Stdio::null()) |
| 399 | .spawn() |
| 400 | .is_ok() |
| 401 | { |
| 402 | return "local-whisper"; |
| 403 | } |
| 404 | } |
| 405 | if std::env::var("GROQ_API_KEY").is_ok_and(|v| !v.trim().is_empty()) { |
| 406 | return "groq"; |
| 407 | } |
| 408 | "provider" |
| 409 | } |
| 410 | |
| 411 | /// Transcribe via local whisper.cpp (free, offline, cross-platform). |
| 412 | /// |
| 413 | /// The whole body is synchronous — temp-file I/O plus `Command::output()`, |
| 414 | /// which blocks for the entire subprocess run — so it runs on the blocking |
| 415 | /// pool rather than a Tokio worker (blocking-call convention, #6149). |
| 416 | pub(crate) async fn transcribe_local_whisper(audio_samples: &[i16]) -> Result<String, String> { |
| 417 | let wav = encode_wav(audio_samples); |
| 418 | tokio::task::spawn_blocking(move || transcribe_local_whisper_blocking(&wav)) |
| 419 | .await |
| 420 | .map_err(|e| e.to_string())? |
| 421 | } |
| 422 | |
| 423 | fn transcribe_local_whisper_blocking(wav: &[u8]) -> Result<String, String> { |
| 424 | // C01-10: a fresh, exclusively created file per call. The old |
| 425 | // `cw-voice-<pid>.wav` name was shared by every transcription in the |
| 426 | // process (an interim pass, the final pass, a runtime-API dictation) and |
| 427 | // predictable in a shared temp dir. Dropping `audio` removes it. |
| 428 | let audio = tempfile::Builder::new() |
| 429 | .prefix("cw-voice-") |
| 430 | .suffix(".wav") |
| 431 | .tempfile() |
| 432 | .map_err(|e| e.to_string())?; |
| 433 | let tmp = audio.path().to_path_buf(); |
| 434 | std::fs::write(&tmp, wav).map_err(|e| e.to_string())?; |
| 435 | // Try each local binary until one succeeds; whisper.cpp outputs to stdout or file. |
| 436 | for bin in LOCAL_WHISPER_BINS { |
| 437 | let output = Command::new(bin) |
| 438 | .arg(tmp.to_string_lossy().as_ref()) |
| 439 | .arg("--model") |
| 440 | .arg("tiny") |
| 441 | .arg("--language") |
| 442 | .arg("auto") |
| 443 | .arg("--output-txt") |
| 444 | .output(); |
| 445 | if let Ok(out) = output |
| 446 | && out.status.success() |
| 447 | { |
| 448 | let txt = String::from_utf8_lossy(&out.stdout).trim().to_string(); |
| 449 | if !txt.is_empty() { |
| 450 | return Ok(txt); |
| 451 | } |
| 452 | // Some builds write to .txt sidecar |
| 453 | let sidecar = tmp.with_extension("txt"); |
| 454 | if let Ok(s) = std::fs::read_to_string(&sidecar) { |
| 455 | let _ = std::fs::remove_file(&sidecar); |
| 456 | if !s.trim().is_empty() { |
| 457 | return Ok(s.trim().to_string()); |
| 458 | } |
| 459 | } |
| 460 | } |
| 461 | } |
| 462 | Err("local whisper not available".into()) |
| 463 | } |
| 464 | |
| 465 | /// Transcribe via Groq Whisper large-v3-turbo (free tier, ~$0.04/hr, fast). |
| 466 | /// Groq is NOT a full CodeWhale provider yet — this is a direct ASR call |
| 467 | /// using `GROQ_API_KEY` only (no provider setup needed). Uses the same |
| 468 | /// chat-completions `input_audio` path as Xiaomi so no `multipart` feature. |
| 469 | pub(crate) async fn transcribe_groq(audio_samples: &[i16]) -> Result<String, String> { |
| 470 | let api_key = std::env::var("GROQ_API_KEY").map_err(|_| "GROQ_API_KEY not set".to_string())?; |
| 471 | let base_url = "https://api.groq.com/openai/v1"; |
| 472 | transcribe_with_model(&api_key, base_url, audio_samples, GROQ_ASR_MODEL, None).await |
| 473 | } |
| 474 | |
| 475 | /// Resolve ASR model/provider preference. |
| 476 | /// Priority: env `CODEWHALE_ASR_MODEL` > auto-detect (local-whisper > groq > |
| 477 | /// provider). Known limitation (C01-10): there is no config-file ASR setting |
| 478 | /// yet — `config` is accepted for the day one exists and is not read, so the |
| 479 | /// env var is the only explicit override. |
| 480 | pub(crate) fn resolve_asr_choice(_config: &Config) -> (String, String) { |
| 481 | let explicit = std::env::var("CODEWHALE_ASR_MODEL") |
| 482 | .ok() |
| 483 | .and_then(|value| explicit_asr_choice(&value)); |
| 484 | // Auto-detect best free: local whisper (offline, no key) > Groq free |
| 485 | // tier > provider ASR (needs key). |
| 486 | let (kind, model) = explicit.unwrap_or_else(|| match detect_free_asr() { |
| 487 | "local-whisper" => ("local-whisper", "tiny"), |
| 488 | "groq" => ("groq", GROQ_ASR_MODEL), |
| 489 | _ => ("provider", ASR_MODEL), |
| 490 | }); |
| 491 | (kind.to_string(), model.to_string()) |
| 492 | } |
| 493 | |
| 494 | /// Map an explicit `CODEWHALE_ASR_MODEL` value to (kind, model). Local |
| 495 | /// spellings are tested first: `local-whisper` and `whisper.cpp` both contain |
| 496 | /// "whisper", which on its own names Groq's hosted Whisper (C01-10). |
| 497 | fn explicit_asr_choice(value: &str) -> Option<(&'static str, &'static str)> { |
| 498 | let m = value.trim().to_ascii_lowercase(); |
| 499 | if m.contains("local") || m.contains("whisper.cpp") || m.contains("whisper-cpp") { |
| 500 | return Some(("local-whisper", "tiny")); |
| 501 | } |
| 502 | if m.contains("groq") || m.contains("whisper") { |
| 503 | return Some(("groq", GROQ_ASR_MODEL)); |
| 504 | } |
| 505 | if m.contains("mimo") || m.contains("xiaomi") { |
| 506 | return Some(("provider", ASR_MODEL)); |
| 507 | } |
| 508 | None |
| 509 | } |
| 510 | |
| 511 | // --- Headless capture (HTTP/native-client path) ---------------------------- |
| 512 | |
| 513 | /// What a headless dictation should do with the finished transcript. |
| 514 | #[derive(Debug, Clone)] |
| 515 | pub enum DictateMode { |
| 516 | /// Transcribe and return the text for insertion into the composer. |
| 517 | Insert, |
| 518 | /// Transcribe, then apply the "send it" / 发送 suffix contract. The |
| 519 | /// outcome's `send` flag tells the client to submit; a bare send |
| 520 | /// instruction yields empty `text` so the client submits its own draft. |
| 521 | Send, |
| 522 | /// AI-assisted dictation that sees the client's composer text — the |
| 523 | /// `/voice-control` pipeline. Only provider ASR can see context; free |
| 524 | /// ASR kinds degrade to plain transcription with `assisted: false`. |
| 525 | Control(String), |
| 526 | } |
| 527 | |
| 528 | /// Machine-readable failure for the headless path so HTTP clients can |
| 529 | /// localize by `reason` rather than parsing message text. |
| 530 | #[derive(Debug)] |
| 531 | pub enum DictateError { |
| 532 | /// No supported recorder binary on this host. |
| 533 | NoRecorder, |
| 534 | /// Recording produced no usable speech segment. |
| 535 | NoSpeech, |
| 536 | /// The selected provider ASR needs an API key that is not configured. |
| 537 | NoProviderAuth, |
| 538 | /// ASR request or transcription failed. |
| 539 | Transcription(String), |
| 540 | } |
| 541 | |
| 542 | impl DictateError { |
| 543 | pub fn reason(&self) -> &'static str { |
| 544 | match self { |
| 545 | Self::NoRecorder => "no_recorder", |
| 546 | Self::NoSpeech => "no_speech", |
| 547 | Self::NoProviderAuth => "no_provider_auth", |
| 548 | Self::Transcription(_) => "transcription_failed", |
| 549 | } |
| 550 | } |
| 551 | } |
| 552 | |
| 553 | impl std::fmt::Display for DictateError { |
| 554 | fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result { |
| 555 | match self { |
| 556 | Self::NoRecorder => write!(f, "no supported voice recorder on this host"), |
| 557 | Self::NoSpeech => write!(f, "no speech detected"), |
| 558 | Self::NoProviderAuth => { |
| 559 | write!(f, "provider ASR requires a configured API key") |
| 560 | } |
| 561 | Self::Transcription(e) => write!(f, "{e}"), |
| 562 | } |
| 563 | } |
| 564 | } |
| 565 | |
| 566 | /// Result of one headless record→transcribe cycle. |
| 567 | #[derive(Debug)] |
| 568 | pub struct DictationOutcome { |
| 569 | /// Final transcript (send suffix already stripped for `Send` mode). |
| 570 | pub text: String, |
| 571 | /// `Send` mode only: the transcript ended with an explicit send phrase. |
| 572 | pub send: bool, |
| 573 | /// `Control` mode only: the composer context reached the model. False |
| 574 | /// when a free ASR kind handled the audio and never saw the context. |
| 575 | pub assisted: bool, |
| 576 | /// Which ASR backend was selected for this capture. |
| 577 | pub asr_kind: String, |
| 578 | pub asr_model: String, |
| 579 | } |
| 580 | |
| 581 | /// Detected recorder binary name, if any (`"sox"`, `"arecord"`, `"rec"`). |
| 582 | pub fn recorder_command() -> Option<&'static str> { |
| 583 | detect_recorder().map(|r| r.cmd) |
| 584 | } |
| 585 | |
| 586 | /// Resolved ASR selection (`kind`, `model`) for capability reporting. |
| 587 | pub fn asr_choice(config: &Config) -> (String, String) { |
| 588 | resolve_asr_choice(config) |
| 589 | } |
| 590 | |
| 591 | /// One record→transcribe cycle with no UI surface: the HTTP/native-client |
| 592 | /// equivalent of `commands::voice::capture_and_transcribe`. Recording runs on a blocking |
| 593 | /// thread; transcription follows the same ASR dispatch as the TUI — |
| 594 | /// explicit `CODEWHALE_ASR_MODEL` > local whisper > Groq > provider — |
| 595 | /// but resolves the provider key lazily so free ASR kinds work without |
| 596 | /// provider auth. |
| 597 | pub async fn dictate_once( |
| 598 | config: &Config, |
| 599 | mode: DictateMode, |
| 600 | ) -> Result<DictationOutcome, DictateError> { |
| 601 | if !is_available() { |
| 602 | return Err(DictateError::NoRecorder); |
| 603 | } |
| 604 | let (samples, _duration) = tokio::task::spawn_blocking(record_audio) |
| 605 | .await |
| 606 | .ok() |
| 607 | .flatten() |
| 608 | .ok_or(DictateError::NoSpeech)?; |
| 609 | |
| 610 | let (asr_kind, asr_model) = resolve_asr_choice(config); |
| 611 | let assisted = asr_kind == "provider" && matches!(&mode, DictateMode::Control(_)); |
| 612 | let composer = match &mode { |
| 613 | DictateMode::Control(composer) => Some(composer.as_str()), |
| 614 | _ => None, |
| 615 | }; |
| 616 | let text = transcribe_selected(config, &asr_kind, &samples, composer).await?; |
| 617 | |
| 618 | let clean = text.trim().to_string(); |
| 619 | let (text, send) = match mode { |
| 620 | DictateMode::Send => { |
| 621 | let (remainder, wants_send) = split_send_suffix(&clean); |
| 622 | (remainder.to_string(), wants_send) |
| 623 | } |
| 624 | _ => (clean, false), |
| 625 | }; |
| 626 | Ok(DictationOutcome { |
| 627 | text, |
| 628 | send, |
| 629 | assisted, |
| 630 | asr_kind, |
| 631 | asr_model, |
| 632 | }) |
| 633 | } |
| 634 | |
| 635 | /// One selected ASR route for both native/HTTP and TUI capture. A failure |
| 636 | /// stays on that route: it must not upload the recording to another provider. |
| 637 | pub(crate) async fn transcribe_selected( |
| 638 | config: &Config, |
| 639 | kind: &str, |
| 640 | samples: &[i16], |
| 641 | composer: Option<&str>, |
| 642 | ) -> Result<String, DictateError> { |
| 643 | let text = match kind { |
| 644 | "local-whisper" => transcribe_local_whisper(samples).await, |
| 645 | "groq" => transcribe_groq(samples).await, |
| 646 | _ => { |
| 647 | let key = config |
| 648 | .active_route_api_key() |
| 649 | .map_err(|_| DictateError::NoProviderAuth)?; |
| 650 | let base = config.active_route_base_url(); |
| 651 | let vendor = config |
| 652 | .openrouter_vendor() |
| 653 | .map_err(|error| DictateError::Transcription(error.to_string()))?; |
| 654 | match composer { |
| 655 | Some(composer) => { |
| 656 | process_voice_control(&key, &base, samples, composer, vendor.as_deref()).await |
| 657 | } |
| 658 | None => transcribe(&key, &base, samples, vendor.as_deref()).await, |
| 659 | } |
| 660 | } |
| 661 | }; |
| 662 | text.map_err(DictateError::Transcription) |
| 663 | } |
| 664 | |
| 665 | #[cfg(test)] |
| 666 | mod tests { |
| 667 | mod selected_asr; |
| 668 | |
| 669 | use super::*; |
| 670 | |
| 671 | #[tokio::test] |
| 672 | async fn voice_requests_preserve_openrouter_vendor_pin() { |
| 673 | use wiremock::matchers::{method, path}; |
| 674 | use wiremock::{Mock, MockServer, ResponseTemplate}; |
| 675 | |
| 676 | let server = MockServer::start().await; |
| 677 | Mock::given(method("POST")) |
| 678 | .and(path("/v1/chat/completions")) |
| 679 | .respond_with(ResponseTemplate::new(200).set_body_json(serde_json::json!({ |
| 680 | "choices": [{ "message": { "content": "{\"text\":\"hello\"}" } }] |
| 681 | }))) |
| 682 | .expect(3) |
| 683 | .mount(&server) |
| 684 | .await; |
| 685 | let base_url = format!("{}/v1", server.uri()); |
| 686 | transcribe( |
| 687 | "fixture-key", |
| 688 | &base_url, |
| 689 | &[0; 16], |
| 690 | Some("chutes/region-fixture"), |
| 691 | ) |
| 692 | .await |
| 693 | .unwrap(); |
| 694 | process_voice_control( |
| 695 | "fixture-key", |
| 696 | &base_url, |
| 697 | &[0; 16], |
| 698 | "existing text", |
| 699 | Some("chutes/region-fixture"), |
| 700 | ) |
| 701 | .await |
| 702 | .unwrap(); |
| 703 | transcribe_with_model("fixture-key", &base_url, &[0; 16], GROQ_ASR_MODEL, None) |
| 704 | .await |
| 705 | .unwrap(); |
| 706 | |
| 707 | let requests = server.received_requests().await.unwrap(); |
| 708 | assert_eq!(requests.len(), 3); |
| 709 | for request in &requests[..2] { |
| 710 | let body: serde_json::Value = serde_json::from_slice(&request.body).unwrap(); |
| 711 | assert_eq!( |
| 712 | body["provider"], |
| 713 | serde_json::json!({"order": ["chutes/region-fixture"], "allow_fallbacks": false}) |
| 714 | ); |
| 715 | } |
| 716 | let independent: serde_json::Value = serde_json::from_slice(&requests[2].body).unwrap(); |
| 717 | assert!(independent.get("provider").is_none()); |
| 718 | } |
| 719 | |
| 720 | #[test] |
| 721 | fn explicit_local_whisper_is_not_routed_to_groq() { |
| 722 | // C01-10: every local spelling contains "whisper"; none may reach the |
| 723 | // hosted Groq backend (which needs a key and sends audio off-host). |
| 724 | for local in ["local-whisper", "whisper.cpp", "whisper-cpp", " LOCAL "] { |
| 725 | assert_eq!( |
| 726 | explicit_asr_choice(local), |
| 727 | Some(("local-whisper", "tiny")), |
| 728 | "{local}" |
| 729 | ); |
| 730 | } |
| 731 | for hosted in ["groq", "whisper", "whisper-large-v3-turbo"] { |
| 732 | assert_eq!( |
| 733 | explicit_asr_choice(hosted), |
| 734 | Some(("groq", GROQ_ASR_MODEL)), |
| 735 | "{hosted}" |
| 736 | ); |
| 737 | } |
| 738 | assert_eq!(explicit_asr_choice("mimo"), Some(("provider", ASR_MODEL))); |
| 739 | assert_eq!(explicit_asr_choice("something-else"), None); |
| 740 | } |
| 741 | |
| 742 | #[test] |
| 743 | fn wav_encoding_produces_valid_header() { |
| 744 | let samples = vec![0i16; 16000]; // 1 second of silence |
| 745 | let wav = encode_wav(&samples); |
| 746 | assert_eq!(&wav[0..4], b"RIFF"); |
| 747 | assert_eq!(&wav[8..12], b"WAVE"); |
| 748 | assert_eq!(&wav[12..16], b"fmt "); |
| 749 | // data size = 16000 * 2 = 32000 |
| 750 | assert_eq!(&wav[4..8], &(36 + 32000u32).to_le_bytes()); |
| 751 | } |
| 752 | |
| 753 | #[test] |
| 754 | fn wav_encoding_empty_is_minimal() { |
| 755 | let wav = encode_wav(&[]); |
| 756 | assert_eq!(wav.len(), 44); |
| 757 | assert_eq!(&wav[4..8], &36u32.to_le_bytes()); |
| 758 | } |
| 759 | |
| 760 | #[test] |
| 761 | fn send_suffix_detected_and_stripped() { |
| 762 | assert_eq!(split_send_suffix("send it"), ("", true)); |
| 763 | assert_eq!(split_send_suffix("Send It!"), ("", true)); |
| 764 | assert_eq!(split_send_suffix("发送"), ("", true)); |
| 765 | assert_eq!(split_send_suffix("發送。"), ("", true)); |
| 766 | assert_eq!( |
| 767 | split_send_suffix("ship the fix, send it"), |
| 768 | ("ship the fix", true) |
| 769 | ); |
| 770 | assert_eq!( |
| 771 | split_send_suffix("修复这个问题，发送"), |
| 772 | ("修复这个问题", true) |
| 773 | ); |
| 774 | } |
| 775 | |
| 776 | #[test] |
| 777 | fn send_suffix_leaves_plain_text_alone() { |
| 778 | assert_eq!(split_send_suffix("send it now"), ("send it now", false)); |
| 779 | assert_eq!( |
| 780 | split_send_suffix("帮我发送一封邮件"), |
| 781 | ("帮我发送一封邮件", false) |
| 782 | ); |
| 783 | assert_eq!(split_send_suffix("发送邮件"), ("发送邮件", false)); |
| 784 | assert_eq!( |
| 785 | split_send_suffix("resend it to the queue"), |
| 786 | ("resend it to the queue", false) |
| 787 | ); |
| 788 | } |
| 789 | |
| 790 | #[test] |
| 791 | fn recorder_detection_does_not_crash() { |
| 792 | // Just verify the function runs without panicking |
| 793 | let _ = is_available(); |
| 794 | } |
| 795 | } |
| 796 |