-
Notifications
You must be signed in to change notification settings - Fork 80
Expand file tree
/
Copy pathdecoder_smoke.cpp
More file actions
485 lines (440 loc) · 20.6 KB
/
Copy pathdecoder_smoke.cpp
File metadata and controls
485 lines (440 loc) · 20.6 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
// decoder_smoke.cpp - real-model gated end-to-end decoder accuracy test.
//
// Loads a real Parakeet v2 GGUF, runs the full pipeline (load → mel →
// encoder → predictor + joint + TDT decode → result accessor population)
// on samples/jfk.wav, and asserts:
//
// 1. transcribe_run completes OK on the canonical sample.
// 2. transcribe_full_text matches the canonical JFK reference text
// to within edit distance 3.
// 3. transcribe_n_tokens > 0 and the per-token accessors return
// non-sentinel values: ids in range, text non-empty, t0_ms ≤
// t1_ms, p in [0, 1].
// 4. transcribe_n_words > 0 and at least one word has the leading
// space stripped (sanity check on the SentencePiece word
// boundary handling in the result builder).
// 5. transcribe_n_segments == 1 (v1 produces a single segment per
// run).
// 6. Explicit token timestamp request returns TRANSCRIBE_TIMESTAMPS_TOKEN
// (Parakeet TDT produces token-level timestamps from encoder
// frame indices).
// 7. transcribe_get_timings reports non-zero mel + encode + decode.
//
// The reference text is the same JFK quote used by validate.py's exact
// transcript comparison. This smoke keeps a small edit-distance budget
// because its job is public API behavior, not tensor-level numerical
// validation.
//
// Gating: built only when TRANSCRIBE_BUILD_REAL_MODEL_TESTS=ON; at run
// time TRANSCRIBE_PARAKEET_GGUF must point at a v2 GGUF. The test
// exits 77 (CTest "skipped") when the model path is unset or missing.
//
// The reference text was validated against v2; v3's encoder weights are
// different and the resulting text may differ, so the test hard-fails
// on v3 rather than silently mis-validating.
#include "transcribe.h"
#include "wav.h"
#include <sys/stat.h>
#include <cmath>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <string>
#include <vector>
namespace {
int g_failures = 0;
#define CHECK(cond) \
do { \
if (!(cond)) { \
std::fprintf(stderr, "FAIL %s:%d: %s\n", __FILE__, __LINE__, #cond); \
++g_failures; \
} \
} while (0)
#define CHECK_EQ_INT(actual, expected) \
do { \
const long long _a = static_cast<long long>(actual); \
const long long _e = static_cast<long long>(expected); \
if (_a != _e) { \
std::fprintf(stderr, "FAIL %s:%d: %s = %lld, expected %lld\n", __FILE__, __LINE__, #actual, _a, _e); \
++g_failures; \
} \
} while (0)
bool file_exists(const std::string & path) {
struct stat st{};
return ::stat(path.c_str(), &st) == 0;
}
// Levenshtein edit distance between two strings. Used for the WER-ish
// gate ("edit distance ≤ 3" per PLAN.md). Implementation is the
// standard rolling-row DP; quadratic in time, linear in memory.
// jfk.wav's reference text is ~110 characters so this is trivially
// fast at the test gate.
int edit_distance(const std::string & a, const std::string & b) {
const int n = static_cast<int>(a.size());
const int m = static_cast<int>(b.size());
std::vector<int> prev(static_cast<size_t>(m + 1));
std::vector<int> curr(static_cast<size_t>(m + 1));
for (int j = 0; j <= m; ++j) {
prev[static_cast<size_t>(j)] = j;
}
for (int i = 1; i <= n; ++i) {
curr[0] = i;
for (int j = 1; j <= m; ++j) {
const int cost = (a[static_cast<size_t>(i - 1)] == b[static_cast<size_t>(j - 1)]) ? 0 : 1;
const int del = prev[static_cast<size_t>(j)] + 1;
const int ins = curr[static_cast<size_t>(j - 1)] + 1;
const int sub = prev[static_cast<size_t>(j - 1)] + cost;
int best = del < ins ? del : ins;
if (sub < best) {
best = sub;
}
curr[static_cast<size_t>(j)] = best;
}
prev.swap(curr);
}
return prev[static_cast<size_t>(m)];
}
// SentencePiece word-boundary marker U+2581 in UTF-8.
constexpr const char k_sp_marker[] = "\xE2\x96\x81";
// Reference text for the canonical jfk.wav validation sample.
const char * const k_jfk_reference_text =
"And so, my fellow Americans, ask not what your country can do for you, "
"ask what you can do for your country.";
constexpr int k_max_edit_distance = 3;
} // namespace
int main() {
// ---- Resolve env -----------------------------------------------
const char * gguf_env = std::getenv("TRANSCRIBE_PARAKEET_GGUF");
if (gguf_env == nullptr || gguf_env[0] == '\0') {
std::fprintf(stderr,
"decoder_smoke: TRANSCRIBE_PARAKEET_GGUF not set; "
"skipping\n");
return 77;
}
if (!file_exists(gguf_env)) {
std::fprintf(stderr, "decoder_smoke: gguf file not found: %s\n", gguf_env);
return 77;
}
const std::string wav_path = std::string(TRANSCRIBE_TEST_SAMPLES_DIR) + "/jfk.wav";
if (!file_exists(wav_path)) {
std::fprintf(stderr, "decoder_smoke: wav not found: %s\n", wav_path.c_str());
return EXIT_FAILURE;
}
// ---- Load model ------------------------------------------------
transcribe_model_load_params mp;
transcribe_model_load_params_init(&mp);
struct transcribe_model * model = nullptr;
{
const transcribe_status st = transcribe_model_load_file(gguf_env, &mp, &model);
if (st != TRANSCRIBE_OK || model == nullptr) {
std::fprintf(stderr, "decoder_smoke: load failed: %s\n", transcribe_status_string(st));
return EXIT_FAILURE;
}
}
// The reference golden was generated against v2; v3 has different
// weights and produces different text. Fail loudly rather than
// silently mis-validating.
{
const std::string variant = transcribe_model_variant_string(model);
if (variant != "tdt-0.6b-v2") {
std::fprintf(stderr,
"decoder_smoke: golden is for tdt-0.6b-v2 only, "
"got \"%s\"\n",
variant.c_str());
transcribe_model_free(model);
return EXIT_FAILURE;
}
}
const std::string backend = transcribe_model_backend(model);
std::fprintf(stdout, "decoder_smoke: backend=%s\n", backend.c_str());
// ---- Load wav --------------------------------------------------
std::vector<float> pcm;
std::string load_err;
if (!transcribe_cli::load_wav_mono_16k(wav_path, pcm, load_err)) {
std::fprintf(stderr, "decoder_smoke: wav load: %s\n", load_err.c_str());
transcribe_model_free(model);
return EXIT_FAILURE;
}
// ---- Init context + run ----------------------------------------
transcribe_session_params cp;
transcribe_session_params_init(&cp);
struct transcribe_session * ctx = nullptr;
{
const transcribe_status st = transcribe_session_init(model, &cp, &ctx);
if (st != TRANSCRIBE_OK || ctx == nullptr) {
std::fprintf(stderr, "decoder_smoke: ctx init: %s\n", transcribe_status_string(st));
transcribe_model_free(model);
return EXIT_FAILURE;
}
}
// Per the public contract, accessors are safe to call before
// transcribe_run and return safe sentinels. Belt-and-braces
// check this so a regression that pre-populates ghost results
// before the first run is caught.
CHECK_EQ_INT(transcribe_n_segments(ctx), 0);
CHECK_EQ_INT(transcribe_n_words(ctx), 0);
CHECK_EQ_INT(transcribe_n_tokens(ctx), 0);
CHECK(std::strcmp(transcribe_full_text(ctx), "") == 0);
CHECK(transcribe_returned_timestamp_kind(ctx) == TRANSCRIBE_TIMESTAMPS_NONE);
transcribe_run_params rp;
transcribe_run_params_init(&rp);
rp.timestamps = TRANSCRIBE_TIMESTAMPS_TOKEN;
{
const transcribe_status st = transcribe_run(ctx, pcm.data(), static_cast<int>(pcm.size()), &rp);
if (st != TRANSCRIBE_OK) {
std::fprintf(stderr, "decoder_smoke: run: %s\n", transcribe_status_string(st));
transcribe_session_free(ctx);
transcribe_model_free(model);
return EXIT_FAILURE;
}
}
// ---- Top-level result ------------------------------------------
const char * full = transcribe_full_text(ctx);
CHECK(full != nullptr);
const std::string actual = full ? full : "";
std::fprintf(stdout, "decoder_smoke: text=\"%s\"\n", actual.c_str());
const int dist = edit_distance(actual, k_jfk_reference_text);
std::fprintf(stdout, "decoder_smoke: edit_distance=%d (tolerance=%d)\n", dist, k_max_edit_distance);
if (dist > k_max_edit_distance) {
std::fprintf(stderr, "FAIL: text edit distance %d exceeds %d\n", dist, k_max_edit_distance);
std::fprintf(stderr, " reference: %s\n", k_jfk_reference_text);
std::fprintf(stderr, " actual: %s\n", actual.c_str());
++g_failures;
}
CHECK(transcribe_returned_timestamp_kind(ctx) == TRANSCRIBE_TIMESTAMPS_TOKEN);
// ---- Per-token sanity ------------------------------------------
const int n_tokens = transcribe_n_tokens(ctx);
CHECK(n_tokens > 0);
int prev_t0 = -1;
int n_with_p = 0;
for (int i = 0; i < n_tokens; ++i) {
transcribe_token tok;
transcribe_token_init(&tok);
CHECK_EQ_INT(transcribe_get_token(ctx, i, &tok), TRANSCRIBE_OK);
CHECK(tok.id >= 0);
CHECK(tok.text != nullptr);
CHECK(tok.t0_ms >= 0);
CHECK(tok.t1_ms >= tok.t0_ms);
CHECK(tok.t0_ms >= prev_t0); // monotone
CHECK_EQ_INT(tok.seg_index, 0);
CHECK(tok.word_index >= 0);
CHECK(tok.p >= 0.0f && tok.p <= 1.0001f && !std::isnan(tok.p));
if (tok.p > 0.0f) {
++n_with_p;
}
prev_t0 = static_cast<int>(tok.t0_ms);
}
// At least most tokens should carry a positive confidence (the
// entropy-based formula gives ~0 for a uniform distribution and
// ~1 for a confident decode; on jfk.wav nearly every token is
// emitted with very high confidence).
CHECK(n_with_p > n_tokens / 2);
// ---- Per-word sanity -------------------------------------------
const int n_words = transcribe_n_words(ctx);
CHECK(n_words > 0);
// The reference text has 22 word-tokens; the result builder
// splits on the SentencePiece marker. We don't pin the exact
// count because punctuation tokens get attached to whichever
// word they follow, but a 11-second jfk should produce on the
// order of ~20 words.
CHECK(n_words >= 15 && n_words <= 30);
int last_word_t1 = -1;
int n_words_with_text = 0;
for (int i = 0; i < n_words; ++i) {
transcribe_word wrd;
transcribe_word_init(&wrd);
CHECK_EQ_INT(transcribe_get_word(ctx, i, &wrd), TRANSCRIBE_OK);
CHECK(wrd.text != nullptr);
if (wrd.text != nullptr && wrd.text[0] != '\0') {
++n_words_with_text;
}
CHECK_EQ_INT(wrd.seg_index, 0);
CHECK(wrd.first_token >= 0 && wrd.first_token < n_tokens);
CHECK(wrd.n_tokens > 0 && wrd.first_token + wrd.n_tokens <= n_tokens);
CHECK(wrd.t0_ms >= 0);
CHECK(wrd.t1_ms >= wrd.t0_ms);
CHECK(wrd.t0_ms >= last_word_t1 - 200); // allow some inter-word slop
// No leading space on a word — the result builder strips it.
if (wrd.text != nullptr && wrd.text[0] == ' ') {
std::fprintf(stderr, "FAIL: word %d has leading space: \"%s\"\n", i, wrd.text);
++g_failures;
}
// No SentencePiece marker should leak into a word's text.
if (wrd.text != nullptr && std::strstr(wrd.text, k_sp_marker) != nullptr) {
std::fprintf(stderr, "FAIL: word %d contains raw SP marker: \"%s\"\n", i, wrd.text);
++g_failures;
}
last_word_t1 = static_cast<int>(wrd.t1_ms);
}
CHECK(n_words_with_text == n_words);
// ---- Per-segment sanity ----------------------------------------
const int n_segments = transcribe_n_segments(ctx);
CHECK_EQ_INT(n_segments, 1);
transcribe_segment seg0;
transcribe_segment_init(&seg0);
CHECK_EQ_INT(transcribe_get_segment(ctx, 0, &seg0), TRANSCRIBE_OK);
CHECK(seg0.text != nullptr);
CHECK(seg0.text != nullptr && seg0.text[0] != '\0');
CHECK(seg0.t0_ms >= 0);
CHECK(seg0.t1_ms > seg0.t0_ms);
CHECK_EQ_INT(seg0.first_word, 0);
CHECK_EQ_INT(seg0.n_words, n_words);
CHECK_EQ_INT(seg0.first_token, 0);
CHECK_EQ_INT(seg0.n_tokens, n_tokens);
// Segment text should equal full_text.
CHECK(seg0.text != nullptr && std::strcmp(seg0.text, full) == 0);
// Segment time should span the audio.
CHECK(seg0.t1_ms <= 12000); // jfk.wav is ~11 s; allow 1 s slack
// ---- Out-of-bounds accessors leave the caller's struct zero-init.
{
transcribe_token oob;
transcribe_token_init(&oob);
CHECK_EQ_INT(transcribe_get_token(ctx, n_tokens, &oob), TRANSCRIBE_OK);
CHECK(oob.text == nullptr);
CHECK_EQ_INT(oob.id, 0);
CHECK_EQ_INT(transcribe_get_token(ctx, -1, &oob), TRANSCRIBE_OK);
CHECK(oob.text == nullptr);
transcribe_word oob_w;
transcribe_word_init(&oob_w);
CHECK_EQ_INT(transcribe_get_word(ctx, n_words, &oob_w), TRANSCRIBE_OK);
CHECK_EQ_INT(oob_w.n_tokens, 0);
transcribe_segment oob_s;
transcribe_segment_init(&oob_s);
CHECK_EQ_INT(transcribe_get_segment(ctx, 5, &oob_s), TRANSCRIBE_OK);
CHECK(oob_s.text == nullptr);
}
// ---- Timings ---------------------------------------------------
transcribe_timings t;
transcribe_timings_init(&t);
CHECK(transcribe_get_timings(ctx, &t) == TRANSCRIBE_OK);
std::fprintf(stdout, "decoder_smoke: timings load=%.2f mel=%.2f encode=%.2f decode=%.2f\n", t.load_ms, t.mel_ms,
t.encode_ms, t.decode_ms);
CHECK(t.load_ms > 0.0f);
CHECK(t.mel_ms > 0.0f);
CHECK(t.encode_ms > 0.0f);
CHECK(t.decode_ms > 0.0f);
// ---- Timestamp ceiling: re-run with coarser requests -----------
//
// Parakeet advertises max_timestamp_kind = TOKEN. A caller that
// asks for WORD, SEGMENT, or NONE gets a result at exactly that
// granularity, with finer-grained tables elided. This is the
// ceiling contract: the request is an upper bound, not an exact
// match. The default run above already covered AUTO→TOKEN.
{
transcribe_run_params rp2;
transcribe_run_params_init(&rp2);
rp2.timestamps = TRANSCRIBE_TIMESTAMPS_WORD;
const transcribe_status st = transcribe_run(ctx, pcm.data(), static_cast<int>(pcm.size()), &rp2);
CHECK(st == TRANSCRIBE_OK);
CHECK(transcribe_returned_timestamp_kind(ctx) == TRANSCRIBE_TIMESTAMPS_WORD);
// Word table still present with real timings.
const int w_n_words = transcribe_n_words(ctx);
CHECK(w_n_words > 0);
if (w_n_words > 0) {
transcribe_word w0;
transcribe_word_init(&w0);
CHECK_EQ_INT(transcribe_get_word(ctx, 0, &w0), TRANSCRIBE_OK);
CHECK(w0.t1_ms >= w0.t0_ms);
}
// Token table elided at WORD granularity, and every word's
// back-reference into the now-empty token table must be
// zeroed so a caller iterating word_first_token /
// word_n_tokens cannot index into a cleared table.
CHECK_EQ_INT(transcribe_n_tokens(ctx), 0);
for (int i = 0; i < w_n_words; ++i) {
transcribe_word wi;
transcribe_word_init(&wi);
CHECK_EQ_INT(transcribe_get_word(ctx, i, &wi), TRANSCRIBE_OK);
CHECK_EQ_INT(wi.first_token, 0);
CHECK_EQ_INT(wi.n_tokens, 0);
}
// Segment's token back-references also zeroed.
{
transcribe_segment s0;
transcribe_segment_init(&s0);
CHECK_EQ_INT(transcribe_get_segment(ctx, 0, &s0), TRANSCRIBE_OK);
CHECK_EQ_INT(s0.first_token, 0);
CHECK_EQ_INT(s0.n_tokens, 0);
}
// Full text still populated.
const char * full2 = transcribe_full_text(ctx);
CHECK(full2 != nullptr && full2[0] != '\0');
}
{
transcribe_run_params rp2;
transcribe_run_params_init(&rp2);
rp2.timestamps = TRANSCRIBE_TIMESTAMPS_SEGMENT;
const transcribe_status st = transcribe_run(ctx, pcm.data(), static_cast<int>(pcm.size()), &rp2);
CHECK(st == TRANSCRIBE_OK);
CHECK(transcribe_returned_timestamp_kind(ctx) == TRANSCRIBE_TIMESTAMPS_SEGMENT);
CHECK_EQ_INT(transcribe_n_words(ctx), 0);
CHECK_EQ_INT(transcribe_n_tokens(ctx), 0);
CHECK_EQ_INT(transcribe_n_segments(ctx), 1);
// Segment timings are still real at SEGMENT granularity.
{
transcribe_segment s0;
transcribe_segment_init(&s0);
CHECK_EQ_INT(transcribe_get_segment(ctx, 0, &s0), TRANSCRIBE_OK);
CHECK(s0.t1_ms > s0.t0_ms);
CHECK(s0.text != nullptr && s0.text[0] != '\0');
}
}
{
transcribe_run_params rp2;
transcribe_run_params_init(&rp2);
rp2.timestamps = TRANSCRIBE_TIMESTAMPS_NONE;
const transcribe_status st = transcribe_run(ctx, pcm.data(), static_cast<int>(pcm.size()), &rp2);
CHECK(st == TRANSCRIBE_OK);
CHECK(transcribe_returned_timestamp_kind(ctx) == TRANSCRIBE_TIMESTAMPS_NONE);
CHECK_EQ_INT(transcribe_n_words(ctx), 0);
CHECK_EQ_INT(transcribe_n_tokens(ctx), 0);
// Segment survives as the text carrier; its t0/t1 are zeroed.
CHECK_EQ_INT(transcribe_n_segments(ctx), 1);
{
transcribe_segment s0;
transcribe_segment_init(&s0);
CHECK_EQ_INT(transcribe_get_segment(ctx, 0, &s0), TRANSCRIBE_OK);
CHECK_EQ_INT(s0.t0_ms, 0);
CHECK_EQ_INT(s0.t1_ms, 0);
CHECK(s0.text != nullptr && s0.text[0] != '\0');
}
}
// ---- Failed transcribe_run clears the prior result -------------
//
// Parakeet's max_timestamp_kind is TOKEN, so the dispatcher
// cannot reject a valid enum request on ceiling grounds. Use an
// out-of-range enum value cast through int to trip the enum-
// range switch at INVALID_ARG. The result-replacement contract
// (include/transcribe.h "Results") says a failing transcribe_run
// must leave the context in the empty sentinel state, not
// silently re-expose the prior successful run's output.
{
// First, seed the context with a real successful run so
// there is a previous result to clobber.
transcribe_run_params rp_ok;
transcribe_run_params_init(&rp_ok);
rp_ok.timestamps = TRANSCRIBE_TIMESTAMPS_TOKEN;
const transcribe_status st_ok = transcribe_run(ctx, pcm.data(), static_cast<int>(pcm.size()), &rp_ok);
CHECK(st_ok == TRANSCRIBE_OK);
CHECK(transcribe_n_tokens(ctx) > 0);
transcribe_run_params rp_bad;
transcribe_run_params_init(&rp_bad);
rp_bad.timestamps = static_cast<transcribe_timestamp_kind>(9999);
const transcribe_status st_bad = transcribe_run(ctx, pcm.data(), static_cast<int>(pcm.size()), &rp_bad);
CHECK(st_bad == TRANSCRIBE_ERR_INVALID_ARG);
// All accessors return their safe sentinels.
CHECK(std::strcmp(transcribe_full_text(ctx), "") == 0);
CHECK_EQ_INT(transcribe_n_segments(ctx), 0);
CHECK_EQ_INT(transcribe_n_words(ctx), 0);
CHECK_EQ_INT(transcribe_n_tokens(ctx), 0);
CHECK(transcribe_returned_timestamp_kind(ctx) == TRANSCRIBE_TIMESTAMPS_NONE);
}
// ---- Teardown --------------------------------------------------
transcribe_session_free(ctx);
transcribe_model_free(model);
if (g_failures > 0) {
std::fprintf(stderr, "decoder_smoke: %d failures\n", g_failures);
return EXIT_FAILURE;
}
std::fprintf(stdout, "decoder_smoke: ok\n");
return EXIT_SUCCESS;
}