feat(voice): improve Romanian STT — hallucination gate + finetuned model

Gemma 4 cloud audio was infeasible (31b-cloud has no audio; E4B broken
upstream, no deploy host), so improve faster-whisper instead.

- Pin temperature=0.0 to disable the fallback ladder that re-decoded unclear
  audio up to 6x (source of the 16-24s latency outliers); reject hallucinated
  segments via avg_logprob/compression_ratio in the new pure _filter_segments.
- Adopt mikr/whisper-small-ro-cv11 (CT2 int8) via configurable voice.stt_model:
  spike showed WER 24%->10%, numbers fixed at source, +0.33s p50 (in budget).
- Add tools/voice_stt_mine.py (log mining) + tools/voice_stt_spike.py (model
  eval with diacritic scoring) + tests for the gate and miner.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
2026-06-27 18:16:16 +00:00
parent ec23d188ec
commit ce273d14db
9 changed files with 664 additions and 16 deletions

1
.gitignore vendored
View File

@@ -29,3 +29,4 @@ memory.bak/
approved-tasks.json approved-tasks.json
dashboard/status.json dashboard/status.json
tools/anaf-monitor/monitor.log tools/anaf-monitor/monitor.log
models/

View File

@@ -110,7 +110,8 @@
], ],
"user_name": "Marius", "user_name": "Marius",
"default_voice": "F1", "default_voice": "F1",
"auto_leave_minutes": 5 "auto_leave_minutes": 5,
"stt_model": "/home/moltbot/echo-core/models/whisper-small-ro-cv11-int8"
}, },
"paths": { "paths": {
"personality": "personality/", "personality": "personality/",

View File

@@ -49,6 +49,14 @@ VAD_WINDOW_BYTES = PACKET_BYTES * (VAD_WINDOW_MS // PACKET_MS)
VAD_THRESHOLD = 0.5 VAD_THRESHOLD = 0.5
SILENCE_FLUSH_MS = 800 SILENCE_FLUSH_MS = 800
NO_SPEECH_DROP_THRESHOLD = 0.6 NO_SPEECH_DROP_THRESHOLD = 0.6
# Hallucination rejection (no re-decode). faster-whisper's default temperature
# is a fallback ladder [0.0..1.0]; on unclear audio it re-decodes the segment up
# to 6x, which is what produced the 16-24s outliers in voice_stt_log.jsonl
# against a >7s conversational-abort budget. We pin temperature=0.0 (no fallback)
# and instead REJECT low-quality segments using the avg_logprob / compression_ratio
# that faster-whisper already computes per segment — zero extra latency.
AVG_LOGPROB_DROP_THRESHOLD = -1.0 # drop seg if avg_logprob below this
COMPRESSION_RATIO_DROP_THRESHOLD = 2.4 # drop seg if gzip ratio above this (repetition/garbage)
PROJECT_ROOT = Path(__file__).resolve().parent.parent.parent PROJECT_ROOT = Path(__file__).resolve().parent.parent.parent
LOGS_DIR = PROJECT_ROOT / "logs" LOGS_DIR = PROJECT_ROOT / "logs"
@@ -83,19 +91,28 @@ _silero_lock = threading.Lock()
def _get_whisper_model() -> Any: def _get_whisper_model() -> Any:
"""Lazy-load faster-whisper ``small`` int8 with the spike-validated """Lazy-load faster-whisper int8 with the spike-validated ``cpu_threads=4``
``cpu_threads=4`` (see ``tasks/voice-bench-results.md``).""" (see ``tasks/voice-bench-results.md``).
Model is configurable via ``voice.stt_model`` (default ``"small"``). It may be
a faster-whisper model name or a path to a local CT2 dir — e.g. the Romanian
Common-Voice finetune that halved WER and fixed number transcription in the
D1 spike (``tools/voice_stt_spike.py``). Custom paths still load with
``local_files_only=True`` since they live on disk."""
global _whisper_model global _whisper_model
if _whisper_model is not None: if _whisper_model is not None:
return _whisper_model return _whisper_model
with _whisper_lock: with _whisper_lock:
if _whisper_model is not None: if _whisper_model is not None:
return _whisper_model return _whisper_model
from src.config import Config
model_id = Config().get("voice.stt_model", "small") or "small"
from faster_whisper import WhisperModel from faster_whisper import WhisperModel
_whisper_model = WhisperModel( _whisper_model = WhisperModel(
"small", device="cpu", compute_type="int8", cpu_threads=4, model_id, device="cpu", compute_type="int8", cpu_threads=4,
local_files_only=True, local_files_only=True,
) )
log.info("STT model loaded: %s", model_id)
return _whisper_model return _whisper_model
@@ -145,6 +162,38 @@ def _pcm48_stereo_to_16_mono(pcm: bytes) -> np.ndarray:
return np.ascontiguousarray(mono16, dtype=np.float32) return np.ascontiguousarray(mono16, dtype=np.float32)
def _filter_segments(segments: Any) -> tuple[list[str], float]:
"""Keep transcribable segments, drop silence and hallucinations.
Pure + side-effect free (no model, no I/O) so the rejection thresholds are
unit-testable with fake segment objects. A segment is dropped when:
- ``no_speech_prob`` is high (silence/non-speech), OR
- ``avg_logprob`` is below ``AVG_LOGPROB_DROP_THRESHOLD`` (decoder unsure), OR
- ``compression_ratio`` exceeds ``COMPRESSION_RATIO_DROP_THRESHOLD`` (looped/garbage).
The avg_logprob/compression checks replace faster-whisper's temperature-fallback
re-decode (the source of the 16-24s latency outliers) with zero-cost rejection.
Returns ``(kept_text_parts, worst_no_speech_prob)``.
"""
text_parts: list[str] = []
worst_no_speech = 0.0
for seg in segments:
no_sp = float(getattr(seg, "no_speech_prob", 0.0) or 0.0)
if no_sp > worst_no_speech:
worst_no_speech = no_sp
if no_sp > NO_SPEECH_DROP_THRESHOLD:
continue
avg_lp = getattr(seg, "avg_logprob", None)
if avg_lp is not None and float(avg_lp) < AVG_LOGPROB_DROP_THRESHOLD:
continue
comp = getattr(seg, "compression_ratio", None)
if comp is not None and float(comp) > COMPRESSION_RATIO_DROP_THRESHOLD:
continue
seg_text = (getattr(seg, "text", "") or "").strip()
if seg_text:
text_parts.append(seg_text)
return text_parts, worst_no_speech
# ---------- VoiceSession ---------- # ---------- VoiceSession ----------
class VoiceSession: class VoiceSession:
@@ -679,6 +728,7 @@ class EchoVoiceSink(AudioSink):
model = _get_whisper_model() model = _get_whisper_model()
segments, _info = model.transcribe( segments, _info = model.transcribe(
mono16, language="ro", beam_size=5, mono16, language="ro", beam_size=5,
temperature=0.0, # no fallback ladder — reject bad segments instead (see thresholds above)
initial_prompt=( initial_prompt=(
"Conversatie in romana cu asistentul Eco (Echo Core). " "Conversatie in romana cu asistentul Eco (Echo Core). "
"Marius i se adreseaza cu 'Salut, Eco', 'Eco' sau 'Echo Core' " "Marius i se adreseaza cu 'Salut, Eco', 'Eco' sau 'Echo Core' "
@@ -689,20 +739,10 @@ class EchoVoiceSink(AudioSink):
"F1, F2, F3, F4, F5. Exemple: vorbeste cu vocea M5, voce F3, " "F1, F2, F3, F4, F5. Exemple: vorbeste cu vocea M5, voce F3, "
"treci pe vocea F1." "treci pe vocea F1."
), ),
hotwords="Eco Echo Core Marius Bianca", hotwords="Eco Echo Core Marius Bianca Bitcoin",
condition_on_previous_text=False, condition_on_previous_text=False,
) )
text_parts: list[str] = [] text_parts, worst_no_speech = _filter_segments(segments)
worst_no_speech = 0.0
for seg in segments:
no_sp = float(getattr(seg, "no_speech_prob", 0.0) or 0.0)
if no_sp > worst_no_speech:
worst_no_speech = no_sp
if no_sp > NO_SPEECH_DROP_THRESHOLD:
continue
seg_text = (getattr(seg, "text", "") or "").strip()
if seg_text:
text_parts.append(seg_text)
if not text_parts: if not text_parts:
return return
text = " ".join(text_parts).strip() text = " ".join(text_parts).strip()

View File

@@ -51,3 +51,17 @@ Lecții capturate din corectările lui Marius. Citește acest fișier la începu
**Greșeala:** Am editat index.json direct, cu o schemă diferită față de ce produce update_notes_index.py. **Greșeala:** Am editat index.json direct, cu o schemă diferită față de ce produce update_notes_index.py.
**Regula:** Niciodată nu scriei manual în `memory/kb/index.json`. Fluxul corect: (1) creezi fișierul `.md` în `memory/kb/<categorie>/`, (2) rulezi `python3 tools/update_notes_index.py`. Dacă ai nevoie să salvezi o notiță din Facebook/video, folosești `scripts/transcribe_video.sh <URL> <lang> --save-kb` care face totul corect. **Regula:** Niciodată nu scriei manual în `memory/kb/index.json`. Fluxul corect: (1) creezi fișierul `.md` în `memory/kb/<categorie>/`, (2) rulezi `python3 tools/update_notes_index.py`. Dacă ai nevoie să salvezi o notiță din Facebook/video, folosești `scripts/transcribe_video.sh <URL> <lang> --save-kb` care face totul corect.
**Când se aplică:** Orice salvare de notiță în KB (Facebook, YouTube, coaching, insights, orice). Dacă ești tentat să `json.dump` în index.json — stop, rulează scriptul. **Când se aplică:** Orice salvare de notiță în KB (Facebook, YouTube, coaching, insights, orice). Dacă ești tentat să `json.dump` în index.json — stop, rulează scriptul.
## Verifică că modelul/tool-ul numit chiar are capabilitatea ÎNAINTE de a planifica în jurul lui
**Data:** 2026-06-27
**Context:** Marius a cerut să folosesc `gemma4:31b-cloud` (Ollama) pentru decodare audio ca alternativă la Whisper. Am verificat pe pagina oficială Ollama: variantele cloud (31b) suportă doar Text+Image — audio există DOAR pe E2B/E4B (edge, local), iar acela e stricat de o regresie upstream deschisă (issue #16584). Premisa cererii era infezabilă.
**Greșeala (evitată):** Dacă planificam direct integrarea fără să verific pagina modelului, scriam cod de cablare Ollama audio care n-ar fi funcționat niciodată. Search-ul generic spunea „Gemma 4 are audio" — adevărat la nivel de familie, fals pentru modelul cloud specific cerut.
**Regula:** Când userul numește un model/serviciu specific pentru o capabilitate (audio, vision, tool-use, context lung), verifică pagina/docs ACELUI model exact înainte de a planifica. Capabilitățile diferă per variantă (cloud vs edge, sizes). Fetch pagina oficială, nu te baza pe search agregat la nivel de familie.
**Când se aplică:** Orice task care pornește de la „folosește modelul X pentru Y". Confirmă X→Y pe sursa primară înainte de plan mode.
## Corecția post-STT de text e cosmetică dacă consumatorul e un LLM — fixează la sursă (model), nu cu dicționar
**Data:** 2026-06-27
**Context:** Plan inițial pentru curățarea transcrierii Whisper avea 4 piese, inclusiv dicționar de restaurare diacritice + canonicalizare wake-word. Două review-uri independente (/autoplan CEO+Eng) au arătat: textul transcris merge la Claude, care citește română fără diacritice perfect; NU există wake-word gate în cod (`on_segment_done` dispatch necondiționat); singurul consumator precis (`detect_voice_change`) e deja fuzz-hardenat. Un spike a confirmat că modelul RO-finetuned (`mikr/whisper-small-ro-cv11`) înjumătățește WER (24%→10%) și fixează numerele la SURSĂ, +0.33s latență.
**Greșeala (evitată):** Construirea unui strat de corecție hand-curat (întreținere perpetuă, risc de regresie pe cuvinte ambigue) când fix-ul real era un model finetuned cu același cost de inferență.
**Regula:** Înainte de a peticit output-ul unui model ML cu post-procesare rule-based, întreabă: (1) cine e CONSUMATORUL textului? (un LLM tolerează erori; un parser regex nu); (2) există un model finetuned care fixează la sursă cu același cost? Spike-uiește modelul ÎNAINTE de a scrie straturi de corecție. Verifică unde merge output-ul prin cod, nu presupune un gate care „pare" că există.
**Când se aplică:** Orice îmbunătățire de calitate STT/OCR/ML output. Tool de spike: `tools/voice_stt_spike.py`.

View File

@@ -0,0 +1,61 @@
# Voice STT Quality — îmbunătățiri Whisper
Branch: `voice/stt-quality`. Origine: cererea de a folosi Gemma 4 cloud pentru audio
(infezabil — `gemma4:31b-cloud` n-are audio, E4B e stricat upstream, fără host de deploy).
Pivot la îmbunătățirea Whisper, validat prin `/autoplan` (CEO + Eng review).
## Ce s-a livrat
### 1. Gate rejection halucinații (cost zero latență) — `src/voice/pipeline.py`
- `model.transcribe(..., temperature=0.0)`**dezactivează scara de fallback** a faster-whisper.
Codul vechi nu pasa `temperature`, deci folosea implicit `[0.0..1.0]` (6 pași) care re-decoda
segmentul pe audio prost → exact sursa latențelor de 24.4s / 16.7s din `voice_stt_log.jsonl`.
- `_filter_segments()` — funcție pură nouă care dropează segmentele cu `no_speech_prob` mare,
`avg_logprob < -1.0` (decoder nesigur) sau `compression_ratio > 2.4` (buclă/gunoi). Zero
re-decodare. Prinde „Care pune o zana judiciul tugea" / „Acest lucru a fost foarte mult".
- `hotwords += Bitcoin`. `initial_prompt` neatins (evită taxa de latență pe fiecare enunț).
- Teste: `tests/test_voice_pipeline_filter.py` (8 cazuri).
### 2. Unealtă de mining — `tools/voice_stt_mine.py`
- CLI read-only peste `voice_stt_log.jsonl`: frecvențe token, tokeni rari (candidați
hotwords/corecții), candidați diacritice lipsă, rânduri suspecte de halucinație.
- Tolerează rânduri fără `text_corrected` (citește `text`). Teste: `tests/test_voice_stt_mine.py` (13).
### 3. Spike model RO-finetuned (D1) — `tools/voice_stt_spike.py`
Compară modele faster-whisper pe audio RO sintetizat (Supertonic) cu ground-truth diacritizat.
**Rezultat (threads=4, beam=5):**
| Model | p50 | p95 | WER | Diacritice |
|-------|-----|-----|-----|-----------|
| `small` (baseline) | 2.59s | 3.04s | 24.2% | 12/20 |
| **`mikr/whisper-small-ro-cv11`** (CT2 int8) | 2.92s | 3.25s | **10.5%** | **17/20** |
- WER se înjumătățește; diacritice 60%→85%; numere PERFECTE (baseline: „120 si 3 delei"
→ finetuned: „o sută douăzeci și trei de lei"). Cost: +0.33s p50 (în bugetul 1.5-3s).
- Modelul CT2: `~/.cache/echo-ct2/whisper-small-ro-cv11-int8` (234M int8).
### 4. Model STT configurabil — `src/voice/pipeline.py::_get_whisper_model`
- Citește `voice.stt_model` din config (default `"small"`). Adopția finetuned = flip config,
nu cod. **Default rămâne `small`** până la decizia de adopție.
## Cum adopți modelul finetuned (când decizi)
```bash
# config.json → "voice": { ..., "stt_model": "/home/moltbot/.cache/echo-ct2/whisper-small-ro-cv11-int8" }
systemctl --user restart echo-core # reload model
```
Re-rulează spike-ul oricând: `python3 tools/voice_stt_spike.py --models "small,<path>" --threads 4`
## Decizii autoplan respinse (din review)
-`temperature=[0.0..0.6]` fallback → regresie latență pe worst-case. Înlocuit cu rejection.
-`canonicalize_wakeword`**nu există wake gate** în cod (verificat); ar fi spart `detect_voice_change`.
- ❌ Dicționar diacritice pe calea Claude → Claude citește română stâlcită OK; finetuned-ul rezolvă la sursă.
-`correct_vocab` / `src/voice/stt_correct.py` → deferate (21 mostre = anecdotă; mining adună întâi date).
## Note de mediu
- `transformers 5.12.1` instalat în `.venv` pentru conversia CT2 (one-time). A downgradat
`tokenizers` 0.23.1→0.22.2 (faster-whisper încă OK, pin `<1` respectat). Se poate `pip uninstall
transformers` dacă nu mai e nevoie de conversii.
- **Pre-existent, neatins de mine:** `tools/tts.py` modificat necommis sparge 2 teste din
`test_voice_normalize.py` (truncare 200 cuvinte). Confirmat: cu `tts.py` committed, testele trec.
```

View File

@@ -0,0 +1,85 @@
"""Tests for src/voice/pipeline.py::_filter_segments — STT hallucination gate.
The gate replaces faster-whisper's temperature-fallback re-decode (the source of
16-24s latency outliers) with zero-cost segment rejection on no_speech_prob,
avg_logprob, and compression_ratio.
"""
import sys
from dataclasses import dataclass
from pathlib import Path
from typing import Optional
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from src.voice.pipeline import ( # noqa: E402
AVG_LOGPROB_DROP_THRESHOLD,
COMPRESSION_RATIO_DROP_THRESHOLD,
NO_SPEECH_DROP_THRESHOLD,
_filter_segments,
)
@dataclass
class FakeSeg:
text: str = ""
no_speech_prob: float = 0.0
avg_logprob: Optional[float] = 0.0
compression_ratio: Optional[float] = 1.0
def test_keeps_clean_segment():
parts, worst = _filter_segments([FakeSeg(text="salut eco", avg_logprob=-0.3, compression_ratio=1.5)])
assert parts == ["salut eco"]
assert worst == 0.0
def test_drops_high_no_speech():
seg = FakeSeg(text="hmm", no_speech_prob=NO_SPEECH_DROP_THRESHOLD + 0.1)
parts, worst = _filter_segments([seg])
assert parts == []
assert worst == NO_SPEECH_DROP_THRESHOLD + 0.1 # still tracked for logging
def test_drops_low_avg_logprob_hallucination():
# "Care pune o zana judiciul tugea" style: decoder unsure
seg = FakeSeg(text="zana judiciul tugea", avg_logprob=AVG_LOGPROB_DROP_THRESHOLD - 0.5)
parts, _ = _filter_segments([seg])
assert parts == []
def test_drops_high_compression_ratio_loop():
seg = FakeSeg(text="da da da da da", compression_ratio=COMPRESSION_RATIO_DROP_THRESHOLD + 1.0)
parts, _ = _filter_segments([seg])
assert parts == []
def test_keeps_when_metrics_missing():
# Older/edge segments may not expose avg_logprob/compression_ratio
seg = FakeSeg(text="ok", avg_logprob=None, compression_ratio=None)
parts, _ = _filter_segments([seg])
assert parts == ["ok"]
def test_drops_empty_text():
parts, _ = _filter_segments([FakeSeg(text=" ", avg_logprob=-0.2)])
assert parts == []
def test_worst_no_speech_is_max_across_segments():
segs = [
FakeSeg(text="a", no_speech_prob=0.1, avg_logprob=-0.2),
FakeSeg(text="b", no_speech_prob=0.4, avg_logprob=-0.2),
]
parts, worst = _filter_segments(segs)
assert parts == ["a", "b"]
assert worst == 0.4
def test_mixed_keep_and_drop():
segs = [
FakeSeg(text="bun venit", avg_logprob=-0.3),
FakeSeg(text="garbage", avg_logprob=-3.0), # dropped: low logprob
FakeSeg(text="la revedere", avg_logprob=-0.5),
]
parts, _ = _filter_segments(segs)
assert parts == ["bun venit", "la revedere"]

View File

@@ -0,0 +1,100 @@
"""Tests for tools/voice_stt_mine.py — STT log mining helpers.
Pure-function coverage: tokenize, token_frequency, rare_tokens,
missing_diacritic_candidates, suspect_rows, row_text (back-compat with rows
that predate the text_corrected field).
"""
import sys
from pathlib import Path
import pytest
sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
from tools.voice_stt_mine import ( # noqa: E402
missing_diacritic_candidates,
rare_tokens,
row_text,
suspect_rows,
token_frequency,
tokenize,
)
def test_tokenize_lowercases_and_drops_punct():
assert tokenize("Salut, Eco!") == ["salut", "eco"]
def test_tokenize_keeps_diacritics():
assert tokenize("ședință și prețul") == ["ședință", "și", "prețul"]
def test_tokenize_drops_digits():
# M3, numbers etc. are not alphabetic word tokens
assert tokenize("M3 are 120 lei") == ["m", "are", "lei"]
def test_tokenize_empty_and_none():
assert tokenize("") == []
assert tokenize(None) == []
def test_row_text_prefers_raw_text_field():
# Mining always wants raw STT output (the `text` field), even once
# newer rows add `text_corrected`.
assert row_text({"text": "cat", "text_corrected": "cât"}) == "cat"
def test_row_text_missing_field():
assert row_text({}) == ""
def test_token_frequency_counts_across_rows():
rows = [{"text": "eco eco"}, {"text": "Eco salut"}]
freq = token_frequency(rows)
assert freq["eco"] == 3
assert freq["salut"] == 1
def test_rare_tokens_returns_singletons_sorted():
rows = [{"text": "eco eco salut bitcoin"}]
rare = rare_tokens(token_frequency(rows))
assert rare == ["bitcoin", "salut"] # eco appears twice -> excluded
assert "eco" not in rare
def test_missing_diacritic_candidates_flags_ascii_words():
rows = [{"text": "pretul este mare"}, {"text": "ședință corectă"}]
cands = missing_diacritic_candidates(token_frequency(rows), min_len=4)
assert "pretul" in cands
assert "mare" in cands
# words carrying diacritics are NOT restore candidates
assert "ședință" not in cands
assert "corectă" not in cands
def test_missing_diacritic_respects_min_len():
rows = [{"text": "cat de bun"}]
cands = missing_diacritic_candidates(token_frequency(rows), min_len=4)
assert "cat" not in cands # len 3 < 4
assert "bun" not in cands
def test_suspect_rows_flags_high_latency():
rows = [
{"text": "ok", "stt_latency_s": 2.0, "no_speech_prob": 0.0},
{"text": "M3.", "stt_latency_s": 24.4, "no_speech_prob": 0.58},
]
suspects = suspect_rows(rows)
assert len(suspects) == 1
assert suspects[0]["text"] == "M3."
def test_suspect_rows_flags_borderline_no_speech():
rows = [{"text": "x", "stt_latency_s": 1.0, "no_speech_prob": 0.55}]
assert len(suspect_rows(rows)) == 1
def test_suspect_rows_tolerates_missing_fields():
# rows without latency/no_speech must not crash
assert suspect_rows([{"text": "x"}]) == []

166
tools/voice_stt_mine.py Normal file
View File

@@ -0,0 +1,166 @@
#!/usr/bin/env python3
"""Mine logs/voice_stt_log.jsonl for STT correction candidates.
Read-only analysis tool. Surfaces what the always-on STT log has captured so
Marius can decide hotwords, spot recurring mistranscriptions, and judge whether
a model swap (e.g. a Romanian-finetuned Whisper) actually helps.
Pure helpers (tokenize / aggregate) are importable and tested; the CLI just
prints reports. Tolerates rows written before the `text_corrected` field
existed (falls back to `text`).
Usage:
python3 tools/voice_stt_mine.py # full report
python3 tools/voice_stt_mine.py --tokens # token frequency only
python3 tools/voice_stt_mine.py --rare # one-off tokens (candidates)
python3 tools/voice_stt_mine.py --suspect # likely hallucination rows
python3 tools/voice_stt_mine.py --log PATH # custom log path
"""
from __future__ import annotations
import argparse
import json
import re
import sys
from collections import Counter
from pathlib import Path
from typing import Iterable, Iterator
PROJECT_ROOT = Path(__file__).resolve().parent.parent
DEFAULT_LOG = PROJECT_ROOT / "logs" / "voice_stt_log.jsonl"
# Latency above this (s) almost always means the decoder thrashed on unclear
# audio — a strong hallucination signal worth reviewing. Mirrors the >7s
# conversational-abort budget from tasks/voice-bench-results.md.
SUSPECT_LATENCY_S = 7.0
SUSPECT_NO_SPEECH = 0.5
_TOKEN_RE = re.compile(r"[A-Za-zĂÂÎȘȚăâîșț]+", re.UNICODE)
# Romanian diacritic letters; a token with none of these is a diacritic-restore
# candidate worth a human glance (not auto-corrected — see plan D2).
_DIACRITICS = set("ĂÂÎȘȚăâîșț")
def read_log(path: Path) -> list[dict]:
"""Parse the JSONL log; skip malformed lines instead of crashing."""
rows: list[dict] = []
if not path.exists():
return rows
with path.open(encoding="utf-8") as f:
for line in f:
line = line.strip()
if not line:
continue
try:
rows.append(json.loads(line))
except json.JSONDecodeError:
continue
return rows
def row_text(row: dict) -> str:
"""Raw transcript for a row. New rows may add `text_corrected`; mining
always wants the raw STT output, which lives in `text`."""
return (row.get("text") or "").strip()
def tokenize(text: str) -> list[str]:
"""Split into alphabetic word tokens, lowercased. Drops digits/punct."""
return [t.lower() for t in _TOKEN_RE.findall(text or "")]
def token_frequency(rows: Iterable[dict]) -> Counter:
counter: Counter = Counter()
for row in rows:
counter.update(tokenize(row_text(row)))
return counter
def rare_tokens(freq: Counter, max_count: int = 1) -> list[str]:
"""Tokens seen at most `max_count` times — candidate mistranscriptions,
proper nouns to add as hotwords, or code-switch garbage."""
return sorted(t for t, c in freq.items() if c <= max_count)
def missing_diacritic_candidates(freq: Counter, min_len: int = 4) -> list[str]:
"""All-ASCII tokens (no Romanian diacritics) of reasonable length, sorted by
frequency. These are the words a diacritic-restore pass would target — kept
as a review list only (v1 does not auto-restore, per plan D2)."""
out = [
(t, c) for t, c in freq.items()
if len(t) >= min_len and not (set(t) & _DIACRITICS) and t.isalpha()
]
out.sort(key=lambda tc: (-tc[1], tc[0]))
return [t for t, _ in out]
def suspect_rows(rows: Iterable[dict]) -> list[dict]:
"""Rows that look like hallucinations: very high latency or borderline
no_speech_prob that still produced text."""
out = []
for row in rows:
lat = float(row.get("stt_latency_s") or 0.0)
nsp = float(row.get("no_speech_prob") or 0.0)
if lat >= SUSPECT_LATENCY_S or nsp >= SUSPECT_NO_SPEECH:
out.append(row)
return out
def _iter_report(rows: list[dict]) -> Iterator[str]:
freq = token_frequency(rows)
yield f"entries: {len(rows)}"
if rows:
lats = [float(r.get("stt_latency_s") or 0.0) for r in rows]
yield f"latency: mean={sum(lats)/len(lats):.2f}s max={max(lats):.2f}s"
yield ""
yield "== top tokens =="
for tok, cnt in freq.most_common(20):
yield f" {cnt:>3} {tok}"
yield ""
yield "== rare tokens (<=1, candidate corrections / hotwords) =="
rare = rare_tokens(freq)
yield " " + (", ".join(rare) if rare else "(none)")
yield ""
yield "== missing-diacritic candidates (review only) =="
cands = missing_diacritic_candidates(freq)[:30]
yield " " + (", ".join(cands) if cands else "(none)")
yield ""
suspects = suspect_rows(rows)
yield f"== likely-hallucination rows ({len(suspects)}) =="
for r in suspects:
yield (f" lat={float(r.get('stt_latency_s') or 0):.1f}s "
f"nsp={float(r.get('no_speech_prob') or 0):.2f} "
f"{row_text(r)!r}")
def main(argv: list[str] | None = None) -> int:
ap = argparse.ArgumentParser(description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--log", type=Path, default=DEFAULT_LOG, help="path to voice_stt_log.jsonl")
ap.add_argument("--tokens", action="store_true", help="token frequency only")
ap.add_argument("--rare", action="store_true", help="one-off tokens only")
ap.add_argument("--suspect", action="store_true", help="likely-hallucination rows only")
args = ap.parse_args(argv)
rows = read_log(args.log)
if not rows:
print(f"no entries in {args.log}", file=sys.stderr)
return 1
freq = token_frequency(rows)
if args.tokens:
for tok, cnt in freq.most_common():
print(f"{cnt:>4} {tok}")
elif args.rare:
print("\n".join(rare_tokens(freq)))
elif args.suspect:
for r in suspect_rows(rows):
print(f"lat={float(r.get('stt_latency_s') or 0):.1f}s "
f"nsp={float(r.get('no_speech_prob') or 0):.2f} {row_text(r)!r}")
else:
print("\n".join(_iter_report(rows)))
return 0
if __name__ == "__main__":
raise SystemExit(main())

180
tools/voice_stt_spike.py Normal file
View File

@@ -0,0 +1,180 @@
#!/usr/bin/env python3
"""STT model spike — compare faster-whisper models on Romanian diacritic accuracy.
Answers the autoplan D1 question: does a Romanian-finetuned whisper-small beat the
generic `small` on diacritics WITHOUT regressing latency? Synthesizes clean RO
audio via Supertonic (ground-truth text known, with diacritics), runs each model,
and scores diacritic preservation + word error rate + latency.
This is an evaluation harness, not production code. Synthetic TTS audio is a clean
probe for diacritic behaviour specifically — it is NOT a proxy for real-mic acoustic
robustness, so weight latency + diacritics, not absolute WER.
Usage:
python3 tools/voice_stt_spike.py --models small,/home/.../whisper-small-ro-cv11-int8
python3 tools/voice_stt_spike.py --models small,<path> --threads 4 --trials 2
"""
from __future__ import annotations
import argparse
import statistics
import sys
import tempfile
import time
import unicodedata
from pathlib import Path
import httpx
SUPERTONIC_URL = "http://127.0.0.1:7788"
# Ground truth with correct Romanian diacritics (mirrors tools/voice_bench.py).
UTTERANCES_RO: list[tuple[str, str]] = [
("short", "Salut, ce mai faci?"),
("conversational", "Stai puțin să mă gândesc la asta."),
("medium", "Am verificat în calendar și avem ședință cu echipa la trei după-amiază."),
("numbers", "Costul total este o sută douăzeci și trei de lei și cincizeci de bani."),
("question", "Marius, vrei să-ți pun pe agenda de mâine să suni la NOAA?"),
("longer", "Vreau să-mi reamintești diseară să verific dacă scriptul de backup a rulat corect."),
]
_DIACRITICS = set("ăâîșțĂÂÎȘȚ")
def _strip_punct_lower(text: str) -> list[str]:
out = []
for raw in text.split():
w = "".join(c for c in raw if c.isalnum() or c in _DIACRITICS)
if w:
out.append(w.lower())
return out
def _deaccent(s: str) -> str:
# Map RO diacritics to base letters for "ignoring diacritics" comparison.
table = str.maketrans("ăâîșțĂÂÎȘȚ", "aaistAAIST")
s = s.translate(table)
return "".join(c for c in unicodedata.normalize("NFD", s)
if unicodedata.category(c) != "Mn")
def wer(ref: list[str], hyp: list[str]) -> float:
"""Word error rate via Levenshtein on token lists."""
n, m = len(ref), len(hyp)
if n == 0:
return 0.0 if m == 0 else 1.0
dp = list(range(m + 1))
for i in range(1, n + 1):
prev = dp[0]
dp[0] = i
for j in range(1, m + 1):
cur = dp[j]
cost = 0 if ref[i - 1] == hyp[j - 1] else 1
dp[j] = min(dp[j] + 1, dp[j - 1] + 1, prev + cost)
prev = cur
return dp[m] / n
def diacritic_score(ref: str, hyp: str) -> tuple[int, int]:
"""For each ground-truth word that carries a diacritic, did the hypothesis
contain that exact diacritized word? Returns (correct, total)."""
ref_words = _strip_punct_lower(ref)
hyp_words = set(_strip_punct_lower(hyp))
hyp_deaccent = {_deaccent(w) for w in hyp_words}
correct = total = 0
for w in ref_words:
if set(w) & _DIACRITICS:
total += 1
if w in hyp_words:
correct += 1
return correct, total
def synthesize(text: str, out_path: Path) -> float:
import wave
r = httpx.post(f"{SUPERTONIC_URL}/v1/audio/speech",
json={"model": "supertonic-3", "input": text, "voice": "M2",
"response_format": "wav", "lang": "ro"}, timeout=60.0)
r.raise_for_status()
out_path.write_bytes(r.content)
with wave.open(str(out_path), "rb") as wf:
return wf.getnframes() / float(wf.getframerate())
def run_model(model_ref: str, wavs: list[tuple[str, str, str, float]],
threads: int, trials: int) -> dict:
from faster_whisper import WhisperModel
t0 = time.perf_counter()
model = WhisperModel(model_ref, device="cpu", compute_type="int8", cpu_threads=threads)
load_s = time.perf_counter() - t0
rows, lats = [], []
dia_c = dia_t = 0
wers = []
for name, ref_text, wav_path, _dur in wavs:
best_text = ""
for trial in range(trials):
t1 = time.perf_counter()
segments, _ = model.transcribe(wav_path, language="ro", beam_size=5,
temperature=0.0, condition_on_previous_text=False)
text = " ".join(s.text.strip() for s in segments).strip()
lats.append(time.perf_counter() - t1)
if trial == 0:
best_text = text
c, tt = diacritic_score(ref_text, best_text)
dia_c += c
dia_t += tt
w = wer(_strip_punct_lower(ref_text), _strip_punct_lower(best_text))
wers.append(w)
rows.append((name, ref_text, best_text, c, tt, w))
return {
"model": model_ref, "load_s": load_s, "rows": rows,
"p50": statistics.median(lats), "p95": sorted(lats)[max(0, int(0.95*(len(lats)-1)))],
"dia_correct": dia_c, "dia_total": dia_t,
"wer": statistics.mean(wers) if wers else 1.0,
}
def main(argv=None) -> int:
ap = argparse.ArgumentParser(description=__doc__,
formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--models", required=True, help="CSV of model names or CT2 dirs")
ap.add_argument("--threads", type=int, default=4)
ap.add_argument("--trials", type=int, default=2)
args = ap.parse_args(argv)
work = Path(tempfile.mkdtemp(prefix="stt_spike_"))
print(f"[spike] synth dir {work}", flush=True)
wavs = []
for name, text in UTTERANCES_RO:
p = work / f"{name}.wav"
dur = synthesize(text, p)
wavs.append((name, text, str(p), dur))
print(f"[spike] TTS {name}: {dur:.2f}s", flush=True)
results = []
for ref in args.models.split(","):
ref = ref.strip()
if not ref:
continue
print(f"[spike] running {ref} (threads={args.threads})…", flush=True)
results.append(run_model(ref, wavs, args.threads, args.trials))
print("\n" + "=" * 72)
print(f"{'model':<42} {'p50':>6} {'p95':>6} {'WER':>6} {'diacr':>8}")
print("-" * 72)
for r in results:
dia = f"{r['dia_correct']}/{r['dia_total']}"
label = Path(r["model"]).name if "/" in r["model"] else r["model"]
print(f"{label:<42} {r['p50']:>5.2f}s {r['p95']:>5.2f}s {r['wer']*100:>5.1f}% {dia:>8}")
print("=" * 72)
for r in results:
label = Path(r["model"]).name if "/" in r["model"] else r["model"]
print(f"\n### {label}")
for name, ref_text, hyp, c, tt, w in r["rows"]:
print(f" [{name}] ref: {ref_text}")
print(f" [{name}] hyp: {hyp} (diacr {c}/{tt}, wer {w*100:.0f}%)")
return 0
if __name__ == "__main__":
sys.exit(main())