RadioWhisper-v1: Domain-Distilled Archival Broadcast ASR
- Trained & Evaluated on GPU:
Tesla T4 - Training Corpus:
117windows across54episodes (15distinct radio series) - Blind Held-Out Eval Corpus:
51windows across23unseen episodes (11distinct radio series) - Training Epoch Cross-Entropy Losses:
[0.3389, 0.177] - Exported CTranslate2 (
float16) Model Size:1547.7 MB
Blind Held-Out Audio Evaluation (Before vs. After Fine-Tuning)
| Model | faster-whisper (CTranslate2 FP16) WER |
Raw HF generate() WER |
Blind Eval Cross-Entropy Loss |
|---|---|---|---|
Stock openai/whisper-large-v3-turbo (Before Training) |
12.14% |
17.39% |
0.4839 |
RadioWhisper-v1 (After Super-Teacher Fine-Tuning) |
8.99% |
12.21% |
0.3277 |
Multi-Series Blind Evaluation Breakdown
| Radio Series | Eval Words | Stock faster-whisper WER |
RadioWhisper-v1 faster-whisper WER |
Stock HF generate() WER |
RadioWhisper-v1 HF generate() WER |
|---|---|---|---|---|---|
| A Little Bit of Everything | 31 |
122.58% |
3.23% |
122.58% |
122.58% |
| Calling All Cars | 128 |
4.69% |
3.12% |
5.47% |
3.12% |
| Coca Cola Topnotchers | 156 |
1.92% |
1.92% |
1.92% |
1.92% |
| Graham MacNamee | 112 |
7.14% |
10.71% |
8.04% |
7.14% |
| Inner Sanctum | 46 |
15.22% |
6.52% |
15.22% |
13.04% |
| New Year's Shows Collection | 44 |
15.91% |
0.0% |
97.73% |
15.91% |
| Sam and Henry | 150 |
29.33% |
25.33% |
32.67% |
28.0% |
| Sherlock Holmes | 293 |
17.06% |
5.8% |
36.52% |
13.65% |
| The Lone Ranger | 160 |
3.75% |
2.5% |
5.0% |
3.75% |
| The Shadow | 855 |
6.67% |
5.03% |
8.42% |
7.95% |
| The Witch's Tale | 785 |
13.89% |
15.67% |
17.45% |
14.65% |
Broadcast Year (1926–1948 Pre-LP Transcription Era) Blind Evaluation Breakdown
| Broadcast Year | Eval Words | Stock faster-whisper WER |
RadioWhisper-v1 faster-whisper WER |
Stock HF generate() WER |
RadioWhisper-v1 HF generate() WER |
|---|---|---|---|---|---|
1926 |
150 |
29.33% |
25.33% |
32.67% |
28.0% |
1927 |
112 |
7.14% |
10.71% |
8.04% |
7.14% |
1928 |
31 |
122.58% |
3.23% |
122.58% |
122.58% |
1929 |
44 |
15.91% |
0.0% |
97.73% |
15.91% |
1930 |
156 |
1.92% |
1.92% |
1.92% |
1.92% |
1931 |
124 |
19.35% |
20.16% |
19.35% |
19.35% |
1932 |
55 |
18.18% |
16.36% |
12.73% |
12.73% |
1933 |
117 |
5.13% |
7.69% |
5.13% |
6.84% |
1934 |
51 |
13.73% |
19.61% |
21.57% |
15.69% |
1935 |
157 |
9.55% |
6.37% |
12.1% |
9.55% |
1936 |
160 |
3.75% |
2.5% |
5.0% |
3.75% |
1937 |
192 |
18.75% |
18.23% |
19.27% |
18.23% |
1938 |
89 |
12.36% |
28.09% |
37.08% |
20.22% |
1939 |
128 |
4.69% |
3.12% |
5.47% |
3.12% |
1940 |
126 |
34.92% |
11.11% |
80.16% |
29.37% |
1941 |
55 |
20.0% |
5.45% |
16.36% |
16.36% |
1942 |
196 |
6.12% |
6.63% |
8.16% |
7.65% |
1943 |
81 |
17.28% |
7.41% |
20.99% |
25.93% |
1944 |
46 |
15.22% |
6.52% |
15.22% |
13.04% |
1945 |
171 |
7.02% |
6.43% |
9.36% |
7.02% |
1946 |
165 |
4.24% |
3.03% |
6.06% |
4.85% |
1947 |
187 |
0.53% |
2.67% |
2.14% |
1.6% |
1948 |
167 |
3.59% |
1.8% |
3.59% |
1.8% |
Sample Blind Evaluation Transcriptions (After Training)
- Sam's Speech at the Lodge Hall (
20.0s):- Ground Truth:
I didn't have no money, then the next best thing to do is to get us something, and I gonna send us something pretty, too. Is you gonna spend a whole ten dollars for a birthday present? I oughta spend more than that, Henry, but I ain't got it, and if I don't send her a birthday present, one of them boys down their in Birmingham is liable to beat my time. We just got about ten more minutes before the whistle blows to go back to work. Let's watch. - RadioWhisper-v1:
I didn't have no money, then the next best thing to do is to get us something, and I'm going to send us something pretty, too. Is you going to spend the whole $10 for a birthday present? I ought to spend more than that, Henry, but I ain't got it, and if I don't send a birthday present, one of them boys down their in Birmingham is liable to beat my time. We just got about 10 more minutes before the whistle blows, so we're back to work. Let's watch.
- Ground Truth:
- Sam's Speech at the Lodge Hall (
45.0s):- Ground Truth:
Just a minute, just a minute. What's your name? You're the fella I'm looking for, all right. You're working on this building here, aren't you? You're the fella, all right. Now, the Indefinite Life Insurance Company of Cincinnati, Ohio sent me here to see you. Who sent you here, mister? - RadioWhisper-v1:
We ain't got much time. Just a minute, just a minute. What's your name? Is you talking to me, mister? My name is Sam Smith. You're the fellow I'm looking for, all right. You're working on this building here, aren't you? Yes, sir, mister. He's working on the building here. You're the fellow, all right. Now, the indefinite life insurance Company of Cincinnati, Ohio, sent me here to see you. Who sent you here, mister?
- Ground Truth:
- Lindbergh and Calvin Coolidge (
250.0s):- Ground Truth:
Climbing on to anything an inch high to three feet high, from a piece of wood to a chair. See if they can't get the first glimpse. Photographers, of course, as usual, are all over the lot. Thousands of them up here. - RadioWhisper-v1:
Climbing onto anything an inch high to three feet high from a piece of wood to a chair. See if they can't get the first glimpse. Photographers, of course, as usual, are all over the lot. Thousands of them up here.
- Ground Truth:
- Lindbergh and Calvin Coolidge (
410.0s):- Ground Truth:
Morning, just three weeks ago yesterday, this wholesome, earnest, fearless, courageous product of America rose into the air from Long Island on a monochrome person, the spirit of St. Louis, in honor of his home and his supporters. It was no - RadioWhisper-v1:
morning just three weeks ago yesterday, this wholesome, earnest, fearless, courageous product of America roused into the air from Long Island on a monoclan cursing the spirit of St. Louis in honor of his home and his supporters. It was known
- Ground Truth:
- Downloads last month
- 14
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support